Military science, Law enforcement | Studies, Essays, Thesises » Peter Bovet Emanuel - Exploring Decision Advantages, Improving Speed, Precision and Efficiency in Military Targeting by Applying Artificial Intelligence

Exploring Decision Advantages Improving Speed, Precision and Efficiency in Military Targeting by Applying Artificial Intelligence Peter Bovet Emanuel Academic dissertation in War Studies Department of War Studies Doctoral Thesis SWEDISH DEFENCE UNIVERSITY THESIS SERIES No 1. 2025 Cover: This illustration depicts a water droplet, labeled “AI,” added into a lake. It symbolizes how artificial intelligence is merged with military decision-making thereby representing the theme of this thesis. The image was generated on 20 January 2025 using ChatGPT-4 with DALL·E 3 (OpenAI). Prompt: “A crystal clear water droplet with ‘AI’ in capital letters falling into a calm lake of mathematical equations. The droplet has the letters ‘AI’ clearly visible on its surface, as well as some random equations that are also subtly visible below the water.” Author Peter Bovet Emanuel ORCID 0000-0001-5475-7349 Publisher Swedish Defence University Publication Year 2025 ISBN (Print)

978-91-88975-52-2 ISSN 2004–6871 DOI 10.62061/leiy7136 Peter Bovet Emanuel 2025 The summary chapter of this thesis is licensed under the terms of the creative commons CC-BY 4.0 license Please note that illustrations and attached articles may have separate licenses or copyright restrictions. Swedish Defence University Box 278 05 115 93 Stockholm www.fhsse Printed by Publit AB, Sweden 2025 Abstract This thesis investigates the integration of artificial intelligence (AI) to augment critical decision-making in military targeting processes, with the intention to make a significant empirical contribution to applied research in War Studies. In the context of contemporary warfare, rapid and informed decision-making is imperative. Grounded in Boyd's OODA loop theory (Observe, Orient, Decide, Act) which emphasizes adaptability and timely action in complex, dynamic situations, this research aims to enhance the speed, precision, and consistency of decision-making within joint

targeting by incorporating AI as an intelligent agent capable of perceiving and acting within these conditioned environments. By constructing and applying two AI models designed to augment the dynamic targeting method, the study addresses two distinct problems in contemporary joint targeting, showcasing practical applications of AI in this context. Model 1 enhances decision-making by improving precision and efficiency in sensor allocation. It identifies optimal locations for deploying target engagement radars (TERs) of medium-range surface-to-air missile systems (MSAMS) and enables decision-makers to achieve more efficient sensor deployment as well as more precise intelligence collection tasks. Model 1 can be utilized for predictive analysis of an adversary's missile system disposition in specific geographical areas and supports the "Observe" and "Orient" stages of Boyd's OODA loop. If validated as an independent intelligence source, Model 1 could initiate

direct target engagements. Model 2 addresses a multi-criteria optimization problem involving multiple targets under given constraints. The results suggest that optimization models can incorporate a commander's targeting guidance to effectively integrate a commander's decision policy as a multi-criteria input into the decision-making calculus mathematically. Model 2 supports all four stages of Boyd's OODA loop and assists in synchronizing feasible attack options to achieve desired effects under time and resource limitations. The findings demonstrate that AI augmentation can significantly expand the decision space for military commanders and offers more opportunities to rapidly exploit, adapt and take the initiative with a greater variety of options. The integration of AI facilitates the transition from hierarchical and linear targeting structures to more dynamic and nonlinear concepts and enhances organizational adaptability and effectiveness under dynamic targeting

conditions. This research underscores the transformative potential of AI in military decision-making and challenges the current human-centric paradigm by introducing AI as an intelligent agent and actor capable of solving problems beyond human limits. The project has important implications for both practitioners and researchers. For practitioners, it offers insights into how AI applications can augment joint targeting practices by improving efficiency and effectiveness in military operations. For researchers, it provides perspectives on the role of AI in military decision-making and how this integration affects command and control arrangements as well as joint warfare concepts. The study suggests that defense organizations should prioritize AI integration to maintain strategic advantages in modern warfare and recommends that future research explore the ethical considerations and long-term impacts of AI-augmented warfare, particularly with regard to command authority and the mechanisms

by which forces and assets are directed. In conclusion, this thesis provides empirical evidence of AI's potential to augment critical military decision-making and proposes two applications for integrating AI into joint targeting processes. By addressing the necessity for military organizations to adopt AI technologies, it contributes to the broader discourse on the future of warfare and the evolving relationship between humans and intelligent machines. Keywords: Artificial Intelligence, Military Targeting, Decision-Making, OODA Loop, Intelligent Agent, Joint Targeting Process, Modelling and Simulation. Sammanfattning Denna avhandling undersöker hur artificiell intelligens (AI) kan integreras för att förstärka det kritiska beslutsfattandet i militära targeting-processer och därigenom göra ett betydande empiriskt bidrag till tillämpad forskning inom War Studies. Med utgångspunkt i Boyds OODA loop-teori (Observe, Orient, Decide, Act), syftar denna forskning till att öka

snabbheten, precisionen och tillförlitligheten i beslutsfattandet inom joint targeting genom att införliva AI som en intelligent agent som kan uppfatta och agera i dessa förutsättningar. Genom att bygga och tillämpa två AI-modeller avsedda att stärka den dynamiska targeting-metoden adresserar studien två olika problem inom samtida joint targeting och visar på hur AI kan användas i praktiken inom detta område. Model 1 förbättrar beslutsfattandet genom att höja precisionen och effektiviteten i sensorallokering. Den identifierar optimala platser för att utplacera target engagement radars (TERs) för medelräckviddiga surface-to-air missile systems (MSAMS) och gör det möjligt för beslutsfattare att uppnå mer effektiv sensorutplacering och mer exakt genomförande av underrättelseinhämtning. Model 1 kan användas för förutsägande analyser av motståndarens robotsystem i specifika geografiska. Om modellen valideras som en oberoende underrättelsekälla skulle Model 1

potentiellt kunna initiera direkta målanfall. Model 2 fokuserar på ett multi-criteria optimization-problem som omfattar flera mål under givna begränsningar. Resultaten pekar på att optimeringsmodeller kan integrera en befälhavares targeting-riktlinjer för att matematiskt införliva befälhavares beslutspolicy som ett flerkriterieunderlag i beslutsfattandet. Resultaten visar att AI-förstärkning väsentligt kan utöka beslutsutrymmet för militära befälhavare och erbjuda fler möjligheter att snabbt utnyttja läget, anpassa sig och ta initiativ med ett större utbud av alternativ. Genom att integrera AI underlättas övergången från hierarkiska och linjära targeting-strukturer till mer dynamiska och icke-linjära koncept, vilket stärker organisationens anpassningsförmåga och effektivitet i samband med dynamiska targeting-förhållanden. Projektet har betydande implikationer både för praktiker och forskare. För praktiker visar det hur AI kan förstärka joint

targeting-rutiner genom att öka effektiviteten och ändamålsenligheten i militära operationer. För forskare erbjuder det insikter i AI:s roll i militärt beslutsfattande och hur denna integration påverkar command and control-arrangemang samt koncept för gemensam krigföring. Avhandlingen föreslår att militära organisationer bör prioritera AI-integration för att behålla strategiska fördelar i modern krigföring och rekommenderar att framtida forskning utforskar de etiska aspekterna och de långsiktiga konsekvenserna av AI-förstärkt krigföring, särskilt med avseende på command authority och på de mekanismer genom vilka militära styrkor och resurser leds. Sammanfattningsvis ger denna avhandling empiriska belägg för AI:s förmåga att förstärka kritiskt militärt beslutsfattande och föreslår två tillämpningar för att integrera AI i joint targeting-processer. Genom att belysa vikten av att militära organisationer anammar AI-teknik bidrar arbetet till den

bredare diskussionen om. Nyckelord: Artificiell Intelligens, Targeting, Beslutsfattning, OODA Loop, Intelligent Agent, Joint Targeting Process, Modellering och Simulering. Acknowledgements This thesis represents the culmination of a five-year academic endeavor and marks the completion of my doctoral research on artificial intelligence aimed at augmenting human decision-making. When my research commenced in 2020, I was concurrently finalizing preparations for a major experiment in northern Sweden and concluding a report on the future development of joint warfighting capabilities for the Swedish Armed Forces. By 2022, it became clear that my work was unfolding amid profound global and regional transformations. Russia’s aggression against Ukraine had escalated into a protracted conflict, where novel technologies and operational concepts were being employed for attritional warfare. Meanwhile, Sweden joined NATO, thereby extending both its domestic security guarantees and external

commitments. During this period, artificial intelligence also emerged as a widely accessible tool, catalyzed by the advent of large language models (LLMs) such as ChatGPT. These developments underscored the relevance of my research on AIdriven decision support, highlighting both its potential applications and the ethical and strategic considerations it entails. However, this journey has not been a solitary one. I would like to express my deepest gratitude to my supervisor, Alastair Finlan, whose patience, insightful editorial feedback, and steadfast support have been invaluable throughout the research process. Together with my assistant supervisor, Arash Heydarian Pashakhanlou, their thoughtful critiques and genuine encouragement pushed me to refine my arguments and uphold academic rigor. I am profoundly thankful to the committee members and the discussant at my final seminar for their discerning critiques and constructive guidance. Their invaluable feedback played a crucial role in

shaping and improving this study. I also extend my sincere appreciation to the Swedish Armed Forces for their generous financial support, which enabled me to dedicate myself fully to this research endeavor. Furthermore, I am deeply grateful to Eray Cakici at IBM, whose consistent technical support and timely advice were instrumental throughout this journey. His dedication and focused efforts provided the foundation that made the optimization research achievable. I am equally appreciative of Ella Olsson at SAAB, whose sharp intellect and collaborative spirit fostered an engaging and thought-provoking environment. Her contributions were key to driving intellectual growth and facilitating the exploration of innovative neural network applications. In addition, I wish to extend my gratitude to other individuals at IBM and SAAB who, in various ways, offered their assistance during this research process. Their contributions, though less visible, were nonetheless significant in shaping the

outcomes of this study. On a personal note, I am profoundly thankful to my family and friends for their unwavering encouragement and support throughout this journey. My wife, Sharon, endured my moments of absent-mindedness with grace and patience, and her unshakable belief in my abilities gave me the strength and motivation to navigate challenges and setbacks. Finally, while I have benefited immensely from the guidance, resources, and support provided by those mentioned here, any interpretations, conclusions, or oversights remain my sole responsibility. The views expressed in this thesis do not necessarily reflect the policies or positions of the aforementioned individuals or institutions. Dedication To Joshua and Neahmay the curiosity that guided my academic endeavors serve as a gentle beacon, encouraging you to pursue knowledge in your own distinctive way and in a field that truly calls to you. Contents Chapter 1 - Introduction . 15 Chapter 2 – Literature review . 27

Chapter 3 - Theoretical framework . 67 Chapter 4 – Method . 105 Chapter 5 – Experiment # 1 . 143 Chapter 6 – Experiment # 2 . 187 Chapter 7 – Conclusions . 232 Key terms and definitions . 260 References . 279 List of tables Table 1 A comparison of mission flights. 40 Table 2 Operationalization . 133 Table 3 Overview of the tasks (activities) of Model 1. 179 Table 4 Targets in Model 2. 215 Table 5 Overview of the tasks (activities) of Model 2 . 223 Table 6 OODA Loop . 240 List of figures Figure 1: Boyd's illustration of the OODA loop . 76 Figure 2 The OODA loop depicted as a simple sequential process . 78 Figure 3: NATO joint targeting cycle (JTC) . 83 Figure 4: The JTC's phases and relations to the intelligence products. 85 Figure 5. The F2T2E2A dynamic targeting process 90 Figure 6. An interpretation of the dynamic targeting process 92 Figure 7. Key intelligence support activities 93 Figure 8. Levels of abstractions from experimentation to War

Science 127 Figure 9. The three units of analysis (UoA) 128 Figure 10: Performance comparison between deep learning (DL) and machine learning (ML) . 146 Figure 11: A basic example of how the model could augment decision-making. 149 Figure 12: Explanation to Step 1 of Model 1 . 153 Figure 13: Creating evenly distributed possible vectors . 154 Figure 14: Filtering out vectors . 155 Figure 15: Finding potential TER target candidates . 155 Figure 16: The output of Model 1. 156 Figure 17: A detailed system description of Model 1. 157 Figure 18: The Buk-M3 MSAMS 9K317M (SA-27 GOLLUM) . 159 Figure 19: The method of creating evenly spaced-out antenna grid generation of radar emitters. 160 Figure 20: The decibel attenuation values. 161 Figure 21: The before part; a raw terrain image without the radar signal coverage calculation. 162 Figure 22: The after part; the terrain image with a radio signal coverage overlay . 163 Figure 23 Comparisons between the four batches . 165 Figure 24: Example

of IoU metric . 166 Figure 25: The output from the first simulation - Simlångsdalen Bridge. E 169 Figure 26: A comparison of a Google Earth Pro picture on the top and the model’s outcome at the bottom. 170 Figure 27: Image from simulation . 171 Figure 28: The output from the third simulation. 171 Figure 29: Distribution of attributes from the SMEs three Optimal TER site locations. 174 Figure 30: The Kveitseid Bridge simulation area. 175 Figure 31: Another vector of the Kveitseid Bridge simulation area. 176 Figure 32: A sample after fine-tuning. 177 Figure 33: Alternative use of the outcome - least covered vector . 183 Figure 34 Schematic overview of the optimization process . 190 Figure 35:The effectors (weapons) data set. 201 Figure 36: The targets data set. 203 Figure 37: The compatibility data set . 205 Figure 38 Illustration of multiple targeting directorates. 250 Figure 39 A minimum viable product used for a demonstration. 254 Chapter 1 - Introduction

“The winning side will be the one that's developed the AI-based decision-making that can outpace their adversary” (Stavridis and Ackerman 2024) 1 Introduction This thesis explores the novel use of artificial intelligence (AI) to augment critical decision-making in military targeting operations. Military targeting is a vital process to identify and select specific assets or forces of an adversary to engage with the goal of achieving mission related criteria while minimizing unintended damage. The thesis investigates how the integration of AI can be incorporated into the targeting process to increase speed, precision and consistency in an attempt to pave a way for new warfare concepts. New forms of warfare such as ‘hyperwar’ are suggested to be ‘AI-driven’ and characterized by speed, complexity and automation (Husain, Allen, and Work 2024). The term ‘hyperwar’ was originally coined in a US Naval Institute journal article in 2017 by General (ret.) John Allen and tech

entrepreneur Amir Husain, who stated that this new form of warfare will have “unparalleled speed enabled by automating decision-making” and that this “will become possible by leveraging artificial intelligence and machine cognition”(2017, 2). Former US Navy Admiral James Stavridis and Elliot Ackerman suggest that the one that “developed the AI-based decision-making that can outpace their adversary” will be on the winning side (2024) 2. While these are ideas that are beginning to form new concepts, they are not yet being applied in warfare. However, AI and machine learning (ML) will enable an acceleration of warfare by increasing the pace of decision-making in military operations. This is likely to challenge previous human-centered strategies in situations that requires rapid decision-making. Successful integration of AI into military power could, as the citation suggests, provide significant strategic advantage. However, developing and deploying AI to augment

decision-making in military targeting requires extensive deliberation, encompassing practical, technical and other foundational considerations. It also raises critical ethical and policy questions including the role of humans in decision-making and explainability / trustworthiness in AI. The use of AI is progressively laying the groundwork for novel 1 James Stavridis, Elliot Ackerman, ’Drone Swarms Are About to Change the Balance of Military Power’, 14 March, 2024, WSJ, https://www.wsjcom/tech/drone-swarms-are-about-to-change-the-balance-of-military-power-e091aa6f 2 Ibid. EXPLORING DECISION ADVANT AGES | 15 warfare practices in which the distribution of work between humans and machines may change and perhaps invoke revisions to established military theory. This project contributes to the growing body of literature by addressing how AI can solve two specific challenges in contemporary joint targeting. It identifies gaps in existing research, defines its research question and

outlines the structure of the thesis. The research problem Modern warfare is increasingly complex and characterized by the need for timely decision-making in dynamic environments. The integration of AI offers potential solutions to these challenges, but large questions remain as to how to integrate it. AI, broadly defined as systems capable of performing human-level tasks under varying and unpredictable circumstances with minimal human oversight, holds promise for optimizing decision-making processes. A US Congressional research report in 2018 on AI stated that although “no official government definition of AI exists” it held that AI can be applied to “Act, Decide, Learn and Perceive” (Hoadley and Lucas 2018, 1,3). AI-Expert Stuart Russell provides a complementary definition and describes AI as “machines are intelligent to the extent that their actions can be expected to achieve their objectives”(2019, 9). This distinction underscores the contrast between human

intelligence, driven by intrinsic objectives, and AI, which is goal-directed based on external input from humans. This section uses three perspectives to illustrate the research problem: the growing amount of data / information and increased tempo in warfare; modifications of the roles and responsibilities; and challenges / opportunities for the military. Data growth and increased tempo The data growth and the increased tempo in warfare challenges humans’ cognitive abilities to collect the relevant information, comprehend a pertinent situational awareness to make primed decisions in time and take action (Meerveld et al. 2023; Ayoub and Payne 2016). Compressed time-frames will impact the command structure and human control. It will “accelerate decision-making” and increase the coordination at a system level due to rapidly changing battle spaces (Ekelhof 2018, 63). It may force adjustments of the targeting processes and procedures Especially if the tactical level utilizes

AI-enabled weapons and enhance their abilities to prosecute targets more rapidly, since that will affect key targeting decisions earlier in the targeting process. As Herwin Meerveld et al suggests, “after all, the aim is to outpace the opponent’s OODA-loop[]and AI-based automation can be an important driver of such efficiency gain.” (2023, 3) Nonetheless, the present ‘human-centric 16 | EXPLORING DECISION ADVANTAGES approach’ (Soare, Singh, and Nouwens 2023, 8) of employing AI, commanding forces / assets and focusing concomitantly on how command should be exercised over increasingly capable resources is questioned by researchers (Kott and Alberts 2017; Liu 2019; J. Johnson 2022a) The technical solutions for decision-making and C2 are deeply rooted in a humancentric thinking and as a consequence the targeting structures of today are disadvantaged from leveraging the technically sophisticated sensor- and weapon systems that exists (Boulanin and Verbruggen 2017; Ekelhof

2018). Consequently, Western targeting organizations (use interchangeably from here on with targeting enterprises) are not structured for rapid decision-making, nor are there processes in place to take advantage of the vast amount of data and information floating around or for that matter to integrate control of all domains (land, sea, air, space, and cyberspace) of modern warfare (Ayoub and Payne 2016; Ducheine, Schmitt, and Osinga 2015). Integrating AI into military operations is far from straightforward. The operational environment is becoming more complex necessitating new targeting concepts and not just technical applications. Challenges include defining AI's role, its trustworthiness, ensuring ethical deployment and adapting existing human-centric military structures. These tensions are yet to be solved The importance of addressing these foundational issues cannot be overstated as they are pivotal to the successful application of AI in decision-making and targeting.

Modifications of the roles and responsibilities The introduction of AI is reshaping roles and responsibilities within military organizations. AI's ability to enhance coordination, intelligence and system-level speed has the potential to push military operations to a pace where machine actions surpass human comprehension (J. Johnson 2020c, 199) Hin Yan Liu (2019) argues that AI will become indispensable in issuing commands, forming a critical component of future warfighting concepts. However, these developments raise concerns about maintaining “meaningful human control” as emphasized by Merel Ekelhof (2019, 343). Ensuring that humans retain the ability to oversee and influence decisions made by AI systems is a key ethical and operational challenge. AI integration also challenges long-standing military practices and requires significant organizational and cultural changes. As Eric Schmidt et al note, overcoming “organizational barriers” and fostering “strategic change”

are essential for successful AI adoption (2021, 291). Meanwhile, the rapid pace of technological advancements in the commercial sector is influencing military capabilities, particularly in weapon and sensor systems. While these innovations offer opportunities to enhance decision-making and targeting processes, they also create EXPLORING DECISION ADVANT AGES | 17 tensions within the traditional paradigms of human-dominated warfare. Successfully navigating these challenges will require a balanced approach that combines human creativity with the efficiency of AI systems. For instance, AI could help mitigate human cognitive limitations, augment decision-making and optimize targeting processes. However, these advancements may come with trade-offs, such as reduced human oversight and increased reliance on automation. Challenges and opportunities in contemporary military decisionmaking processes Despite its potential, integrating AI into military decision-making and targeting presents

significant concerns to be resolved. Frameworks such as the OODA loop, the Military Decision-Making Process (MDMP) and the joint targeting cycle (JTC) provide a foundation for automation, making AI applicable for both general military decision-making and “related processes like the intelligence cycle and the targeting cycle.” (Meerveld et al 2023, 2) These frameworks share common ground in their reliance on two important aspects: comprehension and speed. The project uses the notion of comprehension as an ability to understand and arrive at a conclusion that leads to a decision. Speed is an important element in setting the limits of comprehension. The project describes speed as the ability to accommodate tempo pending operational requirements. It embodies the pace at which comprehension and understanding is attained and decisions or actions are made. The faster a system can retrieve data relevant to a problem, make sense of it, make it actionable, and act, the better (all other

things being equal). AI can enhance both comprehension and speed by rapidly processing data and making it actionable. For example, Bonnie Johnson et al. (2023) highlight the potential of AI in dynamic targeting processes, such as the Find-Fix-Track-Target-Engage-Assess (F2T2EA) method. Their analysis identifies AI applications that could improve the "kill chain," (or F2T2E2A) increasing efficiency and operational effectiveness (2023, 157). Furthermore, as Meerveld et al. argue, failing to utilize AI's capabilities would be both “irresponsible and unethical” given its potential to reduce risks for military personnel and civilians (2023, 5–6). At the same time, preserving human judgment remains essential for interpreting complex scenarios, making ethical decisions and addressing unforeseen challenges. Command and control (C2) systems must ensure accountability and responsibility while leveraging AI's capabilities. This balance is crucial for maintaining trust in

AI systems and ensuring their successful integration into joint operations. 18 | EXPLORING DECISION ADVANT AGES The research gap The problem at the heart of this project is how AI can be applied within joint targeting to support the decision-making process. What has been described hitherto are early shoots of a new development. Military applications where AI actually is implemented to perform decision optimization is an area that is in its formative stage. The research gap has been identified in an extensive study on autonomy and AI conducted based on War Studies by researchers for the US Department of Defence (DoD) (US DoD Defense Science Board 2016b). The study suggests improvements of ‘mission critical functions’, including targeting related processes considered to be too slow (2016b, 91). Another early indicator was the US Congressional report on AI in 2018 identifying that AI systems could “provide decision-makers with the ability to quickly assimilate large volumes of

data and suggest actions faster than current command and control tools.” (Hoadley and Lucas 2018, 26) More recent indicators can be found in other governmental research reports. Mojtahed et al point to the “unrealized potential” in “partially automated information management and accelerated decision-making process”(2023, 3) and a Norwegian Air Power study state that it is most important to develop new concepts where AI and autonomy are integral parts (“AI Og Autonome Systemer” 2024, 5). Other researchers have studied how different AI-technique could support the US Navy targeting (B. Johnson et al 2023) by mapping out the potential areas in which different AI-methods and techniques could be applied. Meerveld et al (2023) suggests that automation could support all stages of the joint targeting cycle. Numerous researchers have suggested various general frameworks of how AI could be applied, including Heather Penney who addresses how the US Air Force must develop its

targeting capability for “a peer conflict [against] dynamic and fleeting targets at a scale, scope, and speed that it has not faced since the Cold War, if ever.” (2023, 10) Other research that has relevance to this project include work by Vinay Chamola et al. who have explored methods to increase the trustworthiness in AI systems (2023) and Iqbal Sarker et al. suggesting directions towards automation including multi-aspect rule-based AI methods (2024). However, none of these are actual AI-applications that has been modeled and then verified in simulations. Beyond this, there are specific research articles that attempt to solve similar problems that this project explores in its two experiments. These are all to a certain degree very technical research papers. Nevertheless, the relevant papers are reviewed in Chapter 2, although none of them attempts to research military decision-making and the targeting process alongside of two tangible experiments the way this project does. This

project contributes to War Studies and applied military studies by illustrating how AI can provide solutions to two specific challenges in contemporary joint targeting. The current state of research is insufficiently developed While there are EXPLORING DECISION ADVANT AGES | 19 general frameworks illustrating how AI could be integrated in the joint targeting process and its decision-making, the paucity of tangible applications reveals a research gap. Moreover, targeting has evolved into a joint concept where additional domains increase the complexity. Modern technologies provide military forces with unprecedented access to information via sensors 3 and to an expanded variety of engagement options via effectors 4. Although these assets are elements within joint targeting and could be determinants of the way C2 is arranged, they are not. The existing command-and-control (C2) approach still has a top-down perspective. An effective targeting process may require new perspectives

perceiving the battlespace as a single continuum to be exploited when and wherever opportunities arise. What is lacking is a comprehensive analysis of how AI impacts critical functions such as joint targeting and how the application of AI and machine learning (ML) techniques could transform existing command and control structures. Consequently, research in this area has the potential to drive significant advancements in joint targeting, fostering innovative targeting practices and enabling the emergence of new targeting concepts. Scope of the thesis This thesis investigates the potential of artificial intelligence (AI) to augment dynamic targeting by enhancing critical human decision-making. In an era where the complexities and tempo of modern warfare continue to intensify, the ability of military organizations to execute all elements of the targeting process with increased speed, precision, and consistency is paramount. The study is grounded in the premise that AI, when integrated

into these elements, has the capacity to optimize dynamic targeting practices and improve decisionmaking efficacy. The research focuses on the role of AI for decision support, emphasizing augmentation rather than replacement of human agency, to ensure a balance between technological efficiency and human oversight. By addressing the implications of AI integration in joint targeting, this thesis contributes to the broader academic discourse on the evolving role of advanced technologies in military operations. The project provides two tangible examples of how AI can be applied to augment decision-making within the joint targeting process. Both experiments are managed and framed as use-cases. The use-case approach is suggested to be a “practical technique” that supports a “test-driven design” (Jacobson, Spence, and Kerr 2016, 96). In this thesis, use-cases are well-defined situations that are conditioned as “a 3 The project refers to sensors as resources that can perceive and

collect information and intelligence. Examples are radars, satellites, aircrafts, or any other platform that has the ability to perceive and collect. 4 The project refers to effectors as resources that can be employed for target engagements. Effectors are platforms plus the weapon or munition. 20 | EXPLORING DECISION ADVANTAGES set of circumstances that explains the problem” that the AI-model is to solve.(Ibid 2016, 96). The project utilize a mixed method approach centered around simulation experiments. The approach corresponds to the authoritative study on military power by Stephen Biddle (2010). As with this project, Biddle used simulation experiments to address “soft variables like force employment in a rigorous way” and thereby “compensating for the weaknesses of individual methods taken alone.”(2010, preface). This also anchors the project within War Studies The investigation uses three different but interrelated concepts as a theoretical framework. The first is

Boyd’s strategic theory and his OODA loop (Boyd 2018) The second is the joint targeting process as defined in military doctrines by NATO and the US Department of Defense (NSO 2021; US DoD Joint Publication 2018b). The third is the intelligent agent as defined by Stuart Russell and Peter Norvig (2014). Together, they provide the foundation from which the project explores the research question. The project applies Boyd's theory to position the experiments within the expansive framework of military decision-making and its necessary prerequisites. Subsequently, it uses doctrines to outline the contemporary approach to targeting within a joint force context and provides explicit guidance on what joint targeting involves and the methods by which it is executed, specifically focusing on the dynamic method. Finally, the concept of the intelligent agent is employed to elucidate what the interventions in the two experiments signify. Collectively, these three concepts enhance and support

one another. The thesis operationalizes the theoretical framework by addressing two distinct problems, each being solved using its own specialized AI-technique (the AIapplication or model). A model is defined here as a constructed structure that specifies its functions, performance metrics and the way it processes inputs to generate outputs that achieve specific goals. Models can facilitate experimentation (simulations) with scenarios (use cases) that would otherwise be ‘challenging’, ‘costly’, or ‘undesirable’ to pursue in real-world settings (Williams 2013, 3–4). The first problem pertains to the localization of high-value targets (HVTs) where a model is employed to enhance human decision-making regarding the allocation of intelligence collection resources (sensor allocation). The second problem involves optimizing human decision-making in multiple target engagements within a dynamic environment characterized by limited time and resources. While the former model

integrates deep learning techniques to address its objectives, the latter utilizes mathematical optimization methods. In combination, these two models exemplify how advanced computational AIapproaches can augment human decision-making in solving different problems intimately related to targeting. EXPLORING DECISION ADVANT AGES | 21 Aim and objective The project aims to investigate the integration of artificial intelligence (AI) into military targeting processes to augment critical decision-making to enhance speed, precision and consistency within joint targeting operations. This involves experimentation where AI either imitates and substitutes human tasks or addresses problems in novel ways. The results are expected to indicate novel perspectives on applying AI to joint targeting. Research objectives: (1) To develop and apply two AI models that address specific problems in contemporary joint targeting to demonstrate practical applications of AI in this context. a. b. Model 1:

Enhance decision-making in sensor allocation by improving precision and efficiency, specifically in deploying target engagement radars (TERs) for medium-range surface-to-air missile systems (MSAMS). Model 2: Address a multi-objective optimization problem involving multiple targets under given constraints to integrate a commander's decision policy mathematically into the decisionmaking calculus. (2) To analyze how AI augmentation can expand the decision space for military commanders to facilitate a shift from hierarchical and linear targeting structures to dynamic and non-linear concepts. (3) To assess the implications of AI integration on military decision-making processes, command and control arrangements and joint warfare concepts. (4) To provide empirical evidence and practical insights for practitioners on how AI applications can augment joint targeting practices. (5) To contribute to the academic discourse by exploring the transformative potential of AI in military

contexts and recommending areas for future research that includes ethical considerations and impacts on command authority. 22 | EXPLORING DECISION ADVANTAGES Research question The central research question of this thesis is: how can AI augment human decisionmaking to create decision advantages within dynamic targeting settings? It is intended to enable an exploration of possible avenues and to elicit the opportunities of AI to create novel openings for a human decision-maker. It specifically addresses the dynamic targeting method, which implies a focus on fast-paced, time-sensitive environments. This allows for an in-depth examination without making the project unmanageable. Contribution This research aims to make a significant empirical contribution to War Studies and applied military studies by demonstrating how AI can address two distinct problems in contemporary joint targeting. To a degree it follows Biddle’s highly innovative approach to applying a multi method analysis

in War Studies but arguably to a greater degree with a more specific focus on the military problem. By constructing and applying two models designed to augment the dynamic targeting method, the studies showcase practical applications of AI in this context. Model 1 enhances decision-making by improving precision and efficiency in sensor allocation. By identifying optimal locations for deploying engagement radars (target acquisition) of medium-range surface-to-air missile systems (SAMS), decisionmakers can achieve more efficient sensor deployment and more precise intelligence collection tasks. Moreover, Model 1 can be utilized for predictive analysis of an adversary's SAMS disposition in specific geographical areas. This capability is valuable for creating ‘intelligence estimates’ (US DoD Joint Publication 2018a, 80) of an adversary and directly supports Boyd's Observe and Orient stages. If considered a validated and independent intelligence source, Model 1 could

potentially be used to initiate direct target engagements. Model 2 addresses a multi-objective optimization problem involving multiple targets under given constraints. The results suggest that optimization models can incorporate a commander's targeting guidance to effectively integrate a commander's decision policy as a multi-criteria input into the targeting process mathematically. The results indicates that Model 2 supports all four stages of Boyd's OODA loop (Observe, Orient, Decide, Act). The project has important implications for both practitioners and researchers. For practitioners, it offers insights into how AI applications can augment the joint targeting practices. For researchers, it provides perspectives on the roles of AI in military decision-making and how this integration affects command and control arrangements as well as joint warfare concepts. EXPLORING DECISION ADVANT AGES | 23 Furthermore, the project extends its findings by exposing future

opportunities through a reconnection with current research on context-awareness discussed in the literature review (Chapter 2). These extensions are significant as they indicate how applied AI could become integral to novel targeting concepts and potentially assisting multiple commanders simultaneously via targeting web services. This advancement offers both practitioners and researchers insights into potentially emerging targeting concepts. The external relevance of this project is substantial, particularly in policy-related domains. It holds utility for both military and governmental strategies concerning future warfare, AI research, upcoming investments, and impending technological integrations. Structure of the thesis The thesis is composed of seven chapters that are compiled into four parts: foundation; framework; empirics; and conclusion. The first two chapters are devoted to surveying foundational issues and dimensions of artificial intelligence, decision-making and joint

warfighting concepts. The two subsequent chapters narrows the focus by setting up the theoretical framework and explaining the methodology and methods applied to the investigation. The thesis moves forward to the two empirical chapters with each engaging in a specific problem derived from contemporary targeting challenges. The final part concludes the empirical results and discusses extensions based upon the results. As the thesis extends over interrelated fields of research overlaps between different sections within the first part (foundation) of the thesis is unavoidable. The occasional overlaps emphasize distinct aspects of shared topics. Chapter 1 introduces the thesis by framing the research problem and highlights the gap in existing scholarship. It establishes the scope, objectives and research question of the project as well as outlining its intended contributions. The chapter also addresses key concerns and provides a brief overview of current research related to the topic.

Chapter 2 presents a comprehensive literature review of key dimensions and issues concerning decision-making, joint targeting and artificial intelligence. It also traces the evolution and convergence of modern targeting and AI from the mid-twentieth century to about a decade ago when these fields began to align. This alignment reflects a perceived shift in the human-technology relationship from a unidirectional flow to a more recursive and bidirectional interaction. Furthermore, it identifies a significant gap in practical applications of AI integration into processes and decisionmaking at higher command levels. 24 | EXPLORING DECISION ADVANT AGES Chapter 3 outlines the theoretical framework that underpins the empirical component of the project. This framework comprises three components: first, Boyd's strategic theory focusing on his conception of the OODA loop(s) (Observe, Orient, Decide, Act) that provides insights into military decision-making processes; second, Western

doctrines on joint targeting that offer guidelines and principles that inform targeting practices within joint forces; and third, the dynamic targeting method and its related intelligence support activities. Chapter 4 details the research methodology and design of the thesis. Beyond merely describing the mixed methods employed, it delves into the role of experimentation within social science research. This includes conceptualization of models to create representations of real-time problems to facilitate analysis, model construction and application. Finally, the chapter develops an analytical tool to support the understanding of the empirical findings presented in Chapters 5 and 6. Chapter 5 is devoted to Model 1 that features the most recent AI technique (deep learning) using a neural network in combination with other techniques to solve a problem related to target intelligence and sensor allocation. Chapter 6 is devoted to Model 2 that is an optimization model engaged in solving the

renowned weapon-to-target assignment problem in a new way. Both empirical chapters are structured similarly to support the comprehension. Chapter 7 synthesizes the results and implications derived from the two experiments, consolidating the key findings of the research. It discusses how the two models can be employed within a broader context of targeting and explores their potential integration into larger systems in novel ways. Additionally, the chapter engages in a brief evaluation of the theory and methods employed, including a reflective discussion on ethical considerations. It specifically examines the implications of the results in relation to Boyd's OODA loop. The thesis concludes by proposing directions for future research and suggests avenues for further exploration. EXPLORING DECISION ADVANT AGES | 25 26 | EXPLORING DECISION ADVANTAGES Chapter 2 – Literature review ‘The war of today is being fought with new weapons, but so was the war of yesterday and the

day before. Drastic change in weapons has been so persistent in the last hundred years that the presence of that factor might be considered one of the constants of strategy. Only those to whom the study of war is novel permit themselves to be swept away by novel elements in the present war.’ (Brodie 1943, preface) Introduction The quote from strategist Bernard Brodie more than 80 years ago relates to the birth of nuclear weapons. As one of the founding fathers of nuclear deterrence, coining it as the “absolute weapon”, Brodie realized that these new weapons would be the centerpiece of deterrence strategies initiating a strategic race for nuclear dominance for the few states that developed such capabilities (Zellen 2015, 107). Even if this statement is still true, the wars of today have been affected by the introduction of artificial intelligence, not only to enhance sensors and effectors. The ‘novel’ element of the present AI technology has abilities to observe, learn, and

adapt. AI is capable of supporting decision-making and to some extent making decisions. That is what makes it different and why it is important to study. The chapter reviews the literature engaged in AI and targeting. It is an attempt to forge a link between two generally separate bodies of work: AI literature and the targeting literature. By connecting the two perspectives relevant dimensions can be analyzed, interrelated, and synthesized into a foundational fabric supporting this project. The structure and arrangement of the review is simple and intended to allow for both an in-depth presentation on the current research and a broader description of the two bodies of work. It reveals that the two bodies of work have begun to intersect as mutual interests are forging a link. This is occurring because military organizations are showing an increased interest in AI, through a growing realization the research, methods, techniques, and applications can alleviate for current challenges

ahead. There also exists an amplified awareness from vendors, businesses, and defense industries to develop and commercialize AI solutions to military counterparts. EXPLORING DECISION ADVANT AGES | 27 The first part of the review focuses on the current research on decision support systems (DSS), methods and practices where AI is utilized to augment human decision-making. It begins by framing the research on decision support using AI that includes a discussion on two key factors: explainability and ethics, when integrating AI into decision support. Moreover, using AI to augment military decision-making can be sensitive depending on what kind of decision-making is being augmented. Since targeting is an area which has been much debated, not least form an ethical point of view, this is also discussed. Although the project’s focus is not on ethics the research does engage in a sensitive topic and therefore explores to a degree the ethics literature. It moves forward to illustrate the

literature brought to attention by the research question and its two experiments. It shows that a common denominator in the form of speed, or the ability to make faster decision, is a core theme in the targeting literature. The review includes the relevant practical work that correlates with this project. In doing so, it explicitly defines the research gap supporting the two empirical chapters. The second part complements the first part by providing a more complete description of the evolution of targeting in the literature: from being a separate thought, a new domain dominated by air power, to becoming an integrated part of a larger joint concept. It also elaborates on how AI became an integral part of military research and development from being a completely separate field of study. After portraying the two trajectories and how they came to forge into one cycle of innovation for military organizations, the second part reviews how researchers describe the trajectory into the future.

The chapter attempts to clarify academic research from more practical work, even if these at times coincide and become hard to divide. It is an attempt to chart previous work as a whole and identify the position of this project within the War Studies field by bringing AI into the heart of military targeting development. The chapter ends with a concluding section supporting both the understanding of the main findings and the linkage to the next chapter Theoretical framework. 28 | EXPLORING DECISION ADVANT AGES Part I Forging a link – AI boosting decision-making and military capability developments There is a trend today in both civilian and military organizations to seek to integrate AI to support decision-making to improve accuracy, efficiency, and reliability of the decisions and subsequent action. From a purely military perspective Meerveld et al (2023) argue that it would be irresponsible not to use AI to support military decision-making. They state that: “military

decision-making consists of an iterative logical planning method to select the best course of action for a given battlefield situation. It can be conducted at levels ranging from tactical to strategic. Each step in this process lends itself to automation. This does not only hold for the MDMP, but also for related processes like the intelligence cycle and the targeting cycle.” (Meerveld et al 2023, 2) The paucity of known military AI-applications to what Meerveld et al. (2023) suggest, necessitates a focus in other directions. Examples of applied decision support systems (DSS) where AI is augmenting human decision-making can be found, for example, in healthcare. These AI-powered decision support systems provide treatment recommendations and improve diagnostic accuracy. By using a combination of electronic health records and real-time monitoring they support clinical decisions used to diagnose diseases, develop personalized treatment plans, and assist clinicians with decision-making

(Alowais et al. 2023) Other examples of DSS are found in monitoring social media for cybersecurity reasons (Lande, Subach, and Puchkov 2020) and in the production industry to enhance decision-making, process optimization, investment prioritization and skills development (Jackson et al. 2024) These successes are likely to influence the military domains Not least since research within the military context argues in favor of boosting its decisionmaking with AI. Meerveld et al frames this as “Given the limitations of human decision-making, the advantage of (partial) automatization with AI can be found both in the temporal dimension and in decision quality.” (2023, 3) They argue that “Ignoring the capabilities of AI to alleviate the limitations of human cognitive performance in military operations[]would be irresponsible and unethical” (Ibid. 2023, 4–5). EXPLORING DECISION ADVANT AGES | 29 According to the US Defense Advanced Research Projects Agency’s (DARPA) official

webpage 5, it is currently running over forty different AI projects, several of these being, or relating to DSS that are sensitive to context, changes, and consequences (Defense Advanced Research Projects Agency 2018a). The Swedish equivalent for military research (FOI) also engages in similar projects including AI’s use in the intelligence and operations processes (Schubert et al. 2019) and how it can enable “partially automated information management and accelerated decision-making process” (Mojtahed et al. 2023, 4) Using AI for “automation and intelligent decisionmaking” (Sarker et al 2024, 1) in critical systems such as military targeting decisionmaking can be made by different AI-techniques and methods Some of these methods such as deep learning (DL) are considered to be at the core of “today’s Fourth Industrial Revolution (4IR or Industry 4.0)” focusing on “technology-driven automation, smart and intelligent systems” (Sarker 2021, 420). However, Sarker argues

that building a feasible model of deep learning remains challenging “due to the dynamic nature and variations of real-world problems and data.”, and these neural networks tend to be “black box[es]” hampering explainability (Ibid. 2021, 420) Explainability is still perhaps the most significant aspect considered when AI applications are to augment human decision-making. Current considerations of using AI in Decision Support Systems (DSS) The current considerations of using AI for decision support revolve around several aspects, including explainability. The need for transparency in AI-driven decisions has led to the development of what is termed XAI. It aims to make AI models interpretable, allowing users to understand and trust the decisions made by the system (Luotsinen et al. 2019) In critical domains like military, healthcare, and finance, XAI could potentially ensure that decisions made are not only accurate but also justifiable. The path towards the decision needs to be

transparent to a human in order to be explainable for the human. In a recent study on the subject, Chamola et al. considers two main types of XAI, “post-hoc explanation and transparency design” (2023, 13), whereas transparency in design enables human understanding of how the model is built and works. The former of the two or post-hoc explanation refers to a built-in ability where a model can explain how and why it came to a conclusion. Additionally, the more significant and important a decision is perceived to be, the more reason for it to be explainable in the aftermath of it. Any results, good or bad, as well as consequences requires accountability for the actions taken and particularly so in military target engagement. However, most algorithms at work are 5 DARPA, “DARPA Tiles Together a Vision of Mosaic Warfare”,https://www.darpamil/work-with-us/darpa-tiles-together-a-vision- of-mosiac-warfare, accessed 2024-06-01. 30 | EXPLORING DECISION ADVANTAGES often not

understood by the ones responsible and accountable. Yet, these algorithms are ‘taken for granted’, despite that some of these systems might be capable of deception (Chamola et al. 2023, 2) Hence, as Chamola et al assert “that the capabilities of current deep learning algorithms are nowhere near to the maximum potential of any AI. Hence, the amount of trust that is given to any AI as of now is highly unjust and excessive" (2023, 2). They highlight an important and perhaps underestimated aspect of AI even if trust is a multi-faceted term. However, other research areas that gain attention in capability development is context-aware AI. Context-awareness includes abilities to learn and adapt in dynamic environments (Matiuzzi Stocchero et al. 2023) and provides transparent decision-making processes (Sarker et al. 2024) as well as aligns with human values and ethical standards (Russell 2019). In a comprehensive study (Xu et al 2021) on the development and application of AI in

scientific research, they conclude that an increased number of scholars pay attention the “change in the computing paradigm from “offline learning + online reasoning” to “online continuous learning,” and thus give the model “the ability of lifelong learning, just like a human being.” (2021, 17) The ability of AI to learn, adapt, and become more sensitive to changing circumstances is likely to make such AI systems more appealing for military organizations. Nevertheless, rule-based AI has benefits In an in-depth study on the matter, Sarker et al. (2024) highlights the utility of knowledge discovery and rulebased AI modelling in tackling cybersecurity issues (2024, 18) By applying rulebased AI, it can offer opportunities for people to set the rules for the AI-model, to assist or augment decision-making as well as automated applications. According to the same study, rule-based AI modelling “holds much promise for enhancing CI security solutions, mainly when considering

automation, intelligence, and transparency, i.e, trustworthiness in decision-making” (Sarker et al 2024, 18) The aspects laid out by Sarker et al. in regard to rule-based AI, especially multiaspect rule-based AI, can enhance transparency As it interrelates to explainability and thereby to the degree of human control or human oversight of AI systems, this can support the development of principles in AI ethics when designing and applying military DSS. Ethical aspects might constraint organizations from incorporating AI in their current non-AI-powered DSS. As AI systems become more integrated into decision-making processes, ethical considerations and governing frameworks will likely gain importance. Organizations and military alliances such as the US Department of Defense and NATO are developing guidelines to ensure that AI systems align with ethical principles and do not perpetuate biases (J. Johnson 2022b; Russell et al 2015; Oniani et al. 2023) In the case of the US Department of

Defence, they have adopted five ethical principles already in 2020, defined as “responsible, equitable, traceable, reliable, and governable” AI, which prompted NATO to release somewhat similar principles, including “lawfulness, responsibility and accountability, explainability and EXPLORING DECISION ADVANT AGES | 31 traceability, reliability, governability, and bias mitigation” (Oniani et al. 2023, 2) This indicates that Western military organizations, at least, are demonstrating a cautious approach to adopting AI and are actively working to mitigate the risks associated with its implementation. The sensitivity of using AI for decision support in a military context The application of AI for decision support in warfare creates tensions and spurs debates. Some of the ongoing discussions are characterized by misconceptions in the narratives of AI, leading to overstated portrayals that add more fiction than science with potential to obscure constructive debates (J. Johnson

2022b; Bode et al 2024; Qiao-Franco and Bode 2023). However, adding AI agents as part of a decisionmaking process remains a sensitive subject The adoption of intelligent machines by military organizations to augment or even make decisions has made researchers become engaged (J. Johnson 2022b, 2020a) From a strategic studies point of view, James Johnson argues that AI-augmentation could mitigate many of the shortcomings inherent to human decision-making, including skewed judgment, cognitive heuristics and groupthink (2019c). His research also challenges this position in other papers by mapping up potential ethical and political problems of having AI in command positions (J. Johnson 2022b) Johnson warns that AI: “cannot effectively, reliably, or safely complement – let alone replace – humans in understanding and apprehending the strategic environment to make predictions and judgments to inform and shape command-and-control (C2) decision-making – the authority and direction

assigned to a commander “ (J. Johnson 2022a, 43). Boulanin et.al argue from a peace and conflict studies perspective (SIPRI study) that the AI-advancements have created many concerns, be it from a legal, ethical, operational or strategic standpoints (Boulanin 2019, I:13). IR researchers Lindsay and Gartzke point out that technological innovation in itself “does not determine military outcomes without the development of complementary doctrines and organizations to employ it” (2020, 6). Other debates concern the perception of human control (Bode and Watts 2021; Verdiesen, Santoni De Sio, and Dignum 2021) and delegations of decision-making authority in relations to autonomous weapon systems (AWS) which vary between states (Scharre 2019; Boulanin and Verbruggen 2017) These debates reveal a problem with differentiation on how to define the crossroad between automated and autonomous weapon systems. 32 | EXPLORING DECISION ADVANTAGES Autonomy has many interpretations, it is

generally understood to be the ability of a machine to perform an intended task without human intervention, using its sensors and computer software to interact with the environment (Boulanin and Verbruggen 2017, 19). Autonomy, in the context of military applications can be integrated in weapon- and sensor systems, but also in other digital systems including command and control systems. The concept of autonomy is very dynamic and possesses scope for ambiguity in terms of rules of engagement. It could potentially transform into an incentive whereby autonomous weapons are perceived as an increasingly attractive asymmetric tool (J. Johnson 2019a) From a military practitioner’s perspective, Scharre argues that the lack of guiding principles and international regulation as the technology continues to evolve generates multiple discourses about interpretating what autonomy in artificial systems mean and how it should be implemented in practice (Scharre, 2019, p. 359) His broad study on

autonomous weapons deconstructs autonomy and suggests that it involves three distinct concepts: “the type of task the machine is performing; the relationship of the human to the machine when performing that task; and the sophistication of the machine’s decision-making when performing the task.” (2018a, 27) The second dimension indicates that there is no clear distinction even in Scharre’s definition, rather that he foresees distinct levels of independence, or self-governance. Scharre therefore adds ‘degrees’ on autonomy for any given task by order of independence: from “semiautonomous” via “supervised autonomous” to “fully autonomous” (2018a, 28). The now commonly used terms that relates to Scharre’s order of independence are: “human in the loop” for semiautonomous operations; “human on the loop” for supervised autonomous operations; and “human out of the loop” for fully autonomous operations (2018a, 29– 30). Boulanin and Verbruggen (2017) favor

a ‘functional approach’ to autonomy recognizing the human–machine command relationship and the functional ability of a machine. From a military policy perspective, the US Department of Defense published its own guidelines in 2023 that established policy and assigned responsibilities for “developing and using autonomous and semi- autonomous functions in weapon systems, including armed platforms that are remotely operated or operated by onboard personnel.”(US Department of Defense 2023, 1) The efforts to define a unified taxonomy on autonomy and the degree or level of self-governance in AI systems is continuing. It is closely related to other terms and concepts including ethics, accountability, and responsibility. However, it is also noteworthy to introduce bias into the discussion. The concept of bias has been defined by Delgado-Rodriguez and Llorca as “the lack of internal validity or incorrect assessment of the association between an exposure and an effect in the target

population in which the statistic estimated has an expectation that does not equal the true value.” (2004, 635) It has been famously researched by, for instance, Tversky and Kahneman concerning judgement under uncertainty (1974) and more recently by Chamola et al. in relation to trustworthy and explainable AI (2023). The project defines bias as something that could make a decision deviate EXPLORING DECISION ADVANT AGES | 33 from an objective reality, for instance, prejudices, predispositions, or conformity that will affect an outcome. The problem with bias in human decision-making is not new as human decision-makers act amidst information overload, uncertainties and, ideally, with an awareness of the limits of their knowledge. AI specialist Ayoub and psychologist Payne (2016) explore AI’s role in mitigating human limitations, including human biases. They argue that humans, “even those with supposed domainspecific expertise, are notoriously poor in predicting future events in

complex social systems. Under certain conditions, AI can do much better” (Ayoub and Payne 2016, 27). As a consequence, humans can be ill suited to manage novel and complex situations. Ayoub and Payne describe how humans judge risks by “weighting impact over likelihood” use “recollections of events as guide to predictions”, and prime attitudes from “selective use of available information to support existing beliefs and conformity” (2016, 16–22). This is echoing previous research, not least Tversky and Kahneman’s findings suggesting three heuristics that humans employ in making judgments under uncertainty:” (i) representativeness, which is usually employed when people are asked to judge the probability that an object or event A belongs to class or process B; (ii) availability of instances or scenarios, which is often employed when people are asked to assess the frequency of a class or the plausibility of a particular development; and (iii) adjustment from an anchor,

which is usually employed in numerical prediction” (1974, 1131). Equally, AI systems may also suffer from bias. These are often inflicted by human flaws when managing the data used by AI that render these AI-models brittle and unpredictable. Notably, AI preserves the bias inherent in the dataset and its underlying code – as long as these are provided by humans (Ayoub and Payne 2016). This bias is fundamental to the contemporary development of AI systems. According to Ayoub and Payne, this can result in a two-phased bias: from the data itself and from the human. They illustrate mitigations from using data-driven, bias-free AI analysis of the biased data that results in only one block of bias in the process flow If AI collects its own data via sensors and categorizes it during analysis, the susceptibility to any human bias would be further mitigated (2016). Unintentional data poisoning can corrupt data just as effectively as deliberate actions by an adversary (Bovet Emanuel, 2024).

This research points to bias in machine learning, since the artificial system could trap human operators into the machine’s bias and create unreliable collaborations leading towards mistrust (Bovet Emanuel, 2024). Likewise, an overreliance in automated information (automation bias) originating from AI systems, can be exploitation by an adversary. As argued by several researchers (Svenmarck et al. 2018a; Schubert et al 2019; Goyal and Bengio 2020) technology is permissive to “manipulation” and sophisticated deception aimed to disrupt, intervene or imperceptibly changing an adversary’s behaviors (Svenmarck et al. 2018b, 7) Recent research even suggests that AI agents may learn how to lie without a human intent priming this behavior (Roff 2020). 34 | EXPLORING DECISION ADVANTAGES A prevailing challenge to contemporary military decision-making – managing speed A prevailing challenge to contemporary military decision-making is speed. The tempo in which a military force is to

understand a situation and decide / act is increasing due to the integration of more automation and a network of connected assets (Boury-Brisset and Berger 2020). Consequently, the developed processes and methods must be used at a faster pace without reducing the quality in any part of it. The fundamentals in military decision-making have not changed since Boyd articulated his OODA loop. Interestingly, it can still be discerned though often modified in illustrations of the essence of joint-all-domains-command and control (JADC2), as indicated in the US DoD approach 6 : “To this end, the JADC2 Strategy articulates three guiding C2 functions of ‘sense,’ ‘make sense,’ and ‘act,’[]”(Department of Defense 2022, 2). The term 'sense' closely aligns with Boyd’s 'Observe,' while 'make sense' corresponds to 'Orient,' and so on. Regardless of the terminology used, the fundamental cycle of rational decision-making remains unchanged. Within

the broader literature, making sense of data and information is highlighted as crucial to decision-making. Notwithstanding the level of command or the functions or forces that are commanded, decisions are primarily made based on data and information processed into assessments or estimates and then integrated to enhance the appreciation of the situation at hand augmenting subsequent decisions. There are challenges to this. First and foremost, as Clark suggests, the future of war seems to invoke new and unparalleled situations of ‘information overload’ (R. M Clark 2020), where the data is readily available but the ability to comprehend the information becomes too great a challenge. Another issue is selecting or filtering vital information from the less relevant and to facilitate the flow and ‘throughput’ of data in the necessary directions to ensure the right prioritization in terms of what data or information that should be dispatched most rapidly. The term ‘age of information

(AoI)’ is used in this context referring to the ‘freshness’ in information (Yates et al. 2021) According to Yates et al “AoI is an end-to-end metric that can be used to characterize latency in status updating systems and applications.”(2021, 1183). Preventing latency is not only a matter of technical communications design but also a question of information management (Zhou et al. 2021) which can be addressed through distributed optimization algorithms (Yang et al. 2023) Challenges associated with decision-making have led scholars and practitioners into various approaches to the problems, giving rise to new and potentially confusing expressions such as decision-centricity and data-centricity. While the former pertains to the continuation of the concept of network-centric warfare (NCW) (Li et al. 2023) and emphasizes decision superiority, the latter focuses on placing the data 6 Decide is in this JADC2 approach understood as being an integral part of making sense. EXPLORING

DECISION ADVANT AGES | 35 at the core of design and decision-making processes (Rajabi 2012). Antoine Bousquet suggests that the term resides within ‘chaoplexic warfare’, a concept that revolves around decentralization and autonomy through the use of technology (Bousquet 2023). Albeit the obfuscations this may bring forward, these terms and concepts have a common denominator in the form of the relevance and timeliness of data for decision-making processes. Arranging and controlling the flow of data and information, the processing (analyses and syntheses) and the timely integration of it are all parts of the decision-making process. The challenges associated with this are well known in military history. To illustrate how the size of information flow has impacted decision-making, and the human in command, Martin van Creveld’s historical study (1989) offers an example from the U.S war in Vietnam that illustrates this: “The Combined Intelligence Center, headed by General

McChristian and serving the South Vietnamese as well as the U.S forces alone received three million pages of enemy documentation per month, of which some 10 percent proved to be of “intelligence value[.]the Army Strategic Communications Facility at Phu Lam was processing some eight thousand messages[.]every day[.]by the end of 1966 it was processing half a million messages per month, a figure that went on to double itself within the next year”. (1989, 246–47) These figures have multiplied over the years making these challenges in a contemporary setting almost unfathomable as suggested by Kenneth Payne (2018), a psychologist and AI-expert. Time becomes important because of its connection to making sense of information and decision-making, especially when overwhelmed with data. Nevertheless, technology can support these challenges and mitigate human computational limitations. It can reduce the time it takes to reach an understanding and a rational decision. The employment of

advanced technologies during Desert Storm made some renowned researchers (DeLanda 1991; Der Derian 1990; Virilio 2002) consider the implications of these technologies in relation to time. Paul Virilio developed a theory of an ‘accelerated culture’ involving speed, technology, and modernity, which has been interpreted for over thirty years. As contended by the renowned security studies scholar James Der Derian, who developed the idea of speed, argued that Virilio “brought the issue of speed back into political and social theory.” (Der Derian 1990, 306) Much of Virilio’s work on speed and warfare relate to communications technologies and attributed them as “weapons of the fourth front” (Virilio 2002, 47). These technologies and the global reach of aerial and space resources make time and space collapse and provide realtime sense-making from large distances (Virilio 2002, 75–78). Olivier Schmitt (Eds Rynning, Schmitt, and Theussen 2021) state that western perceptions of

time in warfare have three dimensions. First the tactically related perception of time as 36 | EXPLORING DECISION ADVANT AGES ‘timing’ or seizing the opportunities that arise in battle. The second dimension of time is the perception of it as a resource to be managed. Finally, when time is conflated with speed and related to tempo or pace in military operations (2021, 207–8). The need for information and data to be meaningful and for timely decision-making continues to drive innovative ideas in the broader literature, while also shaping certain established norms in its implications. Contemporary policy research by RAND organization (Wong et al. 2020) but also by James Johnson (2020b) both suggest that the compression of time and space impacts human decision-making and challenges perceptions of human control. Time in relation to speed and tempo in warfare impose a new set of cognitive apparatus, as elucidated in experiments by US Marine Corps University using a wargaming

platform to explore how to integrate AI in decision-making (B. Jensen, Cuomo, and Whyte 2018; B M Jensen, Whyte, and Cuomo 2019). These experiments are in line with the guidance given by the US Department of Defense (Hoadley and Lucas 2018) specifically addressing experimentation in its report, and a preceding report by the US Defense Science Board suggesting that “autonomy is expected to provide the greatest value by enabling new missions” and “it is essential that this hands-on experimentation explicitly consider the CONOPs, doctrine, and policy implications for new ways to use new systems.” (2016b, 77) The same report also suggests improvements to “mission critical functions” that were considered to be too slow to effectively counter an adaptive adversary (2016a, 95). Researchers engaged in military force employment emphasize the importance of data and information as a military instrument for looming decision advantages. Many point to the integration of advanced

technologies due to needs originating from keeping the human decision-maker aware, informed and able to make sense prior to timely decisions (Ducheine, Schmitt, and Osinga 2015; Kott and Alberts 2017; Lewin 2019; Hoehn 2021). Security studies scholar Merel Ekelhof points towards dynamic and accelerated decision-making in military targeting processes when discussing autonomous weapons and human control (2018). In particular, Ekelhof addresses AI’s role in intelligence support to targeting “because of the massive increase in (and demand for) intelligence, in both quantity and quality, and because rapidly changing battle spaces demand accelerated decision-making” (2018, 63). Efforts to integrate sensor- and weapons systems with command and control into one consolidated seamless network will have to find ways to fuse accessible data, information and intelligence to create actionable knowledge faster than the adversary. AI may enable new analytical approaches to maximize the insights

derived from substantial amounts of data (big data analytics). In a recent article, Johnson illustrated various DARPA-projects engaged in solving related issues such as the KAIROS-program (Knowledge-directed AI Reasoning Over Schemas) which demonstrates how command and control systems infused with AI might function at EXPLORING DECISION ADVANT AGES | 37 a theoretical level (2020a, 8–11). Another is the Real-time Adversarial Intelligence and Decision-making (RAID), a machine learning algorithm designed to predict the goals and movements of an adversary’s forces five hours into the future (Ibid. 2020a, 8–11). Whilst there are a limited number of applications in use that solve more complex challenges, there are ample amounts that uses autonomy in weapon systems. According to a SIPRI-report, autonomy is in at least 56 military systems “to collect and process various types of information[]that might be of critical relevance from a command-and-control perspective” (Boulanin

and Verbruggen 2017, 8). Given DARPA projects aimed at enhancing decision-making and prediction, as well as AIsupported targeting systems, there are valid reasons to believe the challenges discussed in this review are significant. It may also reflect a more simplistic modelling requirement at a sensor- and weapons level as well as conversely a more complex endeavor to AI-applications at a system of system level. Technology has improved human access to data and information which in turn has increased the speed of war through faster decision cycles, and an appeal for increased integration of domains, ways and means. This primes complexity and makes it increasingly difficult for a human commander to direct and guide the forces and assets engaged in simultaneous or parallel targeting engagements. Creating synergies or taking the initiative and seizing opportunities is presumably even harder. Making the required decisions in time and preferably faster than the opponent is becoming more

problematic, though not less tactically important. However, the speed at which decisions can be made at a tactical level relates not only to the tempo of an operation. It needs to correlate with what Paul Brister refers to as “wartime” and grounded in the idea that tactical speed is important when strategic objectives are limited; however as wartime objectives expand, operational and strategic pace marginalizes speed (2021, 51). The idea Brister makes on ‘wartime,’ becomes relevant in a targeting setting and especially when relating back to what Olivier Schmitt refers to when portraying time as a “resource to be managed” (2021, 207). Two general reflections can be made about these ideas. First, managing time is critical at an operational and strategic level to set the conditions for other sources of power to come into play, be it diplomatic, economical, or political. If time can be harnessed and exploited, it could be perceived as a warfighting pacemaker. Second, as more

tactical assets become automated and connected to a network, a web of targeting resources will be at the disposal of anyone (or anything) that has the delegation of authority for target engagement. This would theoretically generate more opportunities that would increase tactical (and arguably operational) speed. 38 | EXPLORING DECISION ADVANTAGES Narrowing the focus: Managing speed and decision-making within joint targeting Joint targeting emerged as a new element of warfare in the twentieth century. Joint targeting is defined in the US DOD Dictionary of Military and Associated Terms as the “process of selecting and prioritizing targets and matching the appropriate response to them, considering operational requirements and capabilities.”(Chairman of the Joint Chiefs of Staff 2021, 211). It is a complex decision mechanism within joint operations that comprise analytical steps and subsequent decisions. It has developed within the air power domain over time and is now integrated

within the concept of joint operations. The United States defines joint operations as “actions conducted by joint forces and those Service forces employed in specified command relationships with each other, which of themselves, do not establish joint forces.” (Ibid 2021, 119). Additionally, joint operations are conducted by a force composed of “two or more elements, assigned or attached, of two or more Military Departments operating under a single joint force commander” (Ibid 2021, 116). This means that within a joint operations context, joint targeting manages a joint force’s targets and matches these with appropriate responses while considering their priorities. Joint targeting utilizes two methods. One is the deliberate method used for engaging targets that allows for preplanning. The second is the dynamic method used for engaging emerging targets that appear in the battlespace. 7 Creating an effective dynamic targeting process requires identifying and prioritizing

emerging targets, deciding on the courses of action, engaging the targets, and assessing the effects. This requires access to a network of sensors and effectors. Whilst the former is necessary to collect data and assess the effects, the latter is integrated to perform the action. However, processing data and information to generate knowledge and understanding is paramount. Equally important is optimizing the use of available resources and making the most appropriate decisions and actions based on the circumstances. According to Johnson, the dynamic targeting process is inherently affined with having limited knowledge and dealing with uncertainties whilst relying on a variety of advanced technology systems (B. Johnson et al 2023, 156) Hence, applying AI augmentation could be a cornerstone to master a web of targeting resources. Contemporary research also argues that the current targeting process is not fast enough (Boury-Brisset and Berger 2020). It requires more “automation” and

integration of AI to reduce the time it takes to proceed from initial detection to the final assessment (Boury-Brisset and Berger 2020, 1–2). It could be argued that the main method of targeting should actually be the dynamic in view of Boyd’s theory (see Chapter 3) and his ideas on adaptation, variation, tempo as well as the aim “to survive and prosper in a non-linear world dominated by change, novelty and 7 The latter is also the method that is used to frame the experiments (the baseline condition) in this research. EXPLORING DECISION ADVANT AGES | 39 uncertainty” (Osinga 2007, 232). The more flexible method of dynamic targeting may be the preferred and most realistic method in contemporary warfare due to heavy dependence on intelligence for a target’s location. This is supported by the UK Ministry of Defence (MoD) response to a Freedom of Information (FoI) request (UK Ministry of Defense 2016). The request is related to the distribution of the two targeting methods.

According to the UK MoD response, 414 British air strikes were carried out in Iraq and Syria during 2015 and 395 were launched under dynamic targeting procedures while just 19 were pre-planned. Accordingly, 95% of the air strikes were done by utilizing the dynamic method (2016). The figures and numbers reveals a predominance of the dynamic targeting method which indicates that the dynamic method is the favored method. US doctrine documents also emphasize that dynamic targeting “occurs in a much more compressed timeline, requiring special consideration and attention” (US Air Force 2021, 23). If dynamic targeting is becoming more frequent this is likely to put more emphasis on a targeting capability that can provision for target engagement within minutes and a capability to provide intelligence support to this method. The former capability calls for highly mobile airborne platforms that can stay aloft for longer periods of time than regular fighter jets and for high-precision,

hyper-sonic long-range weapons that have a short time from launch to impact. Remotely piloted aircraft systems (RPAS) such as the MQ-9 Reaper are arguably the most versatile and cost-effective alternative of the two 8. Revisiting the same UK MoD’s response, it also encloses a comparison (Table 1) between the number of missions flown by drones (MQ-9 Reaper) and aircrafts (Tornado, Typhoon) respectively (UK Ministry of Defense 2016): Table 1 The table refers to the number of missions flown by Drones (Reaper) and Aircrafts (Tornado and Typhoon) in Iraq and Syria during the period of January, February, and March 2016. Source: UK Ministry of Defense, 2016 A review of the total number of missions flown reveals that 183 missions were performed by MQ-9 Reapers, 198 missions by Tornados and 152 missions by Typhoons. Although the table lacks information concerning the specific mission 8 Given its significant loiter time (24hours mission time) , wide-range sensors, multi-mode communications

suite, and precision weapons, it provides a unique capability to perform strike, coordination, and reconnaissance against high-value, fleeting, and timesensitive targets: Air Force, “MQ-9 Reaper”, https://www.afmil/About-Us/Fact-Sheets/Display/Article/104470/mq-9-reaper/ , accessed 10 Aug 2024. 40 | EXPLORING DECISION ADVANT AGES profiles, this could be compiled from the UK MoD’s webpage as the data is available there to allow for this. However, the employment of RPAS is significant Assuming the two targeting methods from 2015 (95% dynamic targeting) were the same in 2016 (Table 1), the table reveals that the MQ-9 Reaper (first column ‘REAPER’) performed just as many total air strikes as the two types of crewed aircrafts (‘TORNADO’ and ‘TYPHOON’). Although these facts are far from a complete record of how contemporary targeting is being applied, it is still an indication of a shift towards more dynamic targeting executions. Importantly, it also indicates that

weaponized drones like the MQ-9 Reaper are perceived as an important asset for targeting. Current research on military AI applications in targeting There are factors and indicators that suggests a shift towards more dynamic targeting executions. However, moving towards this requires an augmentation of the human ability to comprehend and make timely targeting decisions. It will necessitate “advanced automated support” using AI/ML techniques for “optimized collection planning/tasking and weapon-target assignment[]” (Boury-Brisset and Berger 2020, 1). The time factor in joint targeting decision-making is also directly related to the significance and priority of a target. The common doctrinal definitions are time-sensitive targets (TST), high-value targets (HVT) and component critical target (CCT). TSTs are generally defined by the strategic level, HVTs by the operational level and CCTs by the tactical (component) level (NSO 2021). The significance and priority of a target

correlates to a joint force’s commitment to allocate its forces and assets. Moreover, priority stems from the urgency of striking a particular target. It is a hierarchy of importance based on a combination of aspects including the assessed criticality (how vital it is for the adversary), target significance (how vital it is to meet the joint force’s objectives) and the level of threat it poses to the joint force. Utilizing AI for optimized intelligence collection planning and sensor allocation Previous work on how AI can support decision-making in targeting by optimizing the intelligence collection planning and sensor allocation is limited. This project makes a contribution to this literature. However, more generic research on the matter is available. One of the most significant is by Boury-Brisset and Berger , for example, that discusses the benefits and challenges of artificial intelligence and machine learning (AI/ML) in support of both intelligence and targeting (2020). Their

point of departure is that “the military operational environment is evolving and becoming more complex, involving multiple domains and actors and an accelerated EXPLORING DECISION ADVANT AGES | 41 tempo” (2020, 1). They also acknowledge that AI/ML is progressively used to support different processes, including the “Intelligence, Surveillance and Reconnaissance (ISR), Command and control (Observe/Orient/Decide/Act - OODA loop), or Targeting (e.g Find, Fix, Track, Target, Engage, Assess – F2T2EA processes)”(2020, 2). In a context of augmenting decision-making, Boury-Brisset and Berger suggest that these processes benefit from AI solutions which can support automation and optimization of “competing tasks and resources, and multiple objectives under a diversity of constraints (cost, risk, or communication)[]to derive best collection/fire plans”(2020, 3). However, their proposed solution is a generic framework or “representative AI/ML applications” (2020, 5,9) that

illustrates how various AI/ML techniques could be employed. Nevertheless, this valuable study points to the same areas of improvement (optimizing intelligence collection and weapons to target assignment) as this project does, although their research is tailored towards “classification and fusion techniques for automatic target recognitionin support of enhanced situational awareness and targeting” (2020, 2). The paucity of more tangible research indicates a research gap in a military context within the field of War Studies. However, the techniques and methods intended to be applied in support of the first experiment share common ground in other fields of research. The commonality is mainly related to the method of deep learning (DL) using neural networks and the specific technique of semantic segmentation. Previous research includes comprehensive studies of the potential of deep learning (Sarker 2021; Xu et al. 2021), the limitations (LeCun 2018) including studies on DL’s

vulnerabilities to cyber-attacks (Kamrani et al. 2023), explainability (XAI) in military deep learning applications (Luotsinen et al. 2019) such as wargaming (M Cohen et al. 2020) Current research suggesting the more general utility on deep learning for military applications to improve efficiency in decision-making can also be found in Cai et al. (2023) and in Schubert et al (2018) Specific evaluations of its potentials in targeting practices have been proposed by B. Johnson et al. (2023), exploring “the use of artificial intelligence (AI) for enhancing the naval tactical kill chain” (2023, 155). They underscore the importance of an “effective and appropriate design and engineering of AI-enhanced and/or AI-enabled kill chains[]to achieving tactical superiority” suggesting that their project “provides an analytical foundation[]for continued research” on the different subtasks identified as parts of the targeting process (B. Johnson et al 2023, 164) To date, no known

research in War Studies or related fields has investigated the use of neural networks to understand how topography impacts radiofrequency (RF) signal attenuation through the use of semantic segmentation of satellite images. Notwithstanding this apparent research gap, two popular architectures used for semantic segmentation are U-Net and ResNet. U-Net was first introduced in 2015 by Ronneberger et. al in their paper ‘U-Net: Convolutional Networks for Biomedical Image Segmentation’ (Ronneberger, Fischer, and Brox 2015), and received its name 42 | EXPLORING DECISION ADVANT AGES from the U-shaped encoder-decoder network architecture. U-Net architecture is based on a fully convolutional network that consists of an encoder-decoder structure with skip connections that allow information to flow directly from the encoder to the decoder. According to Ronneberger etal there was until 2015 a large consensus that successful training of deep networks required several thousand annotated

training samples. In their paper, they present a network and training strategy that uses the data samples more efficiently and show that such a network can be trained end-to-end from very few images and outperforms the prior best method (2015, 240–41). Since then, U-Net has also been applied to terrain segmentation in several studies including Biasutti et al. (2019) applying the architecture to high-resolution LiDAR 9 data for terrain feature extraction. Others such as Abdollahi et al. (2022) extract building features from high-resolution imagery to enhance updates in geospatial databases. Wang and Miao (2022) also use automatic detection and building extraction from remote-sensing images for urban planning. ResNet (short for residual network) was first introduced in 2015 by He et al. in the paper ‘Deep Residual Learning for Image Recognition’ (He et al. 2015) The architecture is designed to address the problem of a degradation of accuracy when more layers are added in a

network. Additional layers were thought to increase the accuracy, although the opposite (a degradation) occurred instead. To solve this degradation problem in deep neural networks, He et al. introduced “shortcut connections” that allow information to bypass one or more layers (2015, 771). Their results showed that the residual learning principle was generic, therefore applicable in other vision and non-vision problems. ResNet has been applied to terrain and elevation data in several studies including ‘A Novel Semantic Segmentation Network Based on a Hybrid Framework Combining a Convolutional Neural Network and Transformer for Deep Space Rock Images’ by Fan et al. (2023) using it for rock detection to avoid obstacle and for path planning. It has also been employed for ‘Semantic segmentation of land cover from high resolution multispectral satellite images by spectral-spatial convolutional neural network’ by Saralioglu and Gungor (2022) who used it to extract more accurate

land cover information from very highresolution satellite images. One of the advantages of using these architectures (U-Net or ResNet) is that they enable the employment of a pretrained neural network, which can significantly speed up the training process and may reduce the amount of training data needed to reach a desired performance of the neural network. Overall, U-Net and ResNet are both powerful architectures for semantic segmentation in terrain and elevation data and have been used in studies to extract valuable information about the terrain topography. This research has chosen to have a U-Net architecture utilizing a ResNet34 backbone for emitter location determination. The ResNet34’s extraction 9 Light Detection and Ranging (LiDAR) EXPLORING DECISION ADVANT AGES | 43 capabilities for classification tasks were therefore combined with the spatially aware architecture of U-Net designed for segmentation tasks. Although there has been progress in deep learning over the past

decade, it is still in an exploration phase. Semantic segmentation of high-resolution images, including satellite images, by neural networks has been applied to different problems and appears useful in the various research areas revealed in the review. However, no research has been identified that uses this method within a clearly defined military problem setting. Utilizing AI for optimizing weapon-target assignments This narrow literature review relates specifically to the second experiment and the efforts of optimizing weapon-target assignments. Compared to the use of the novel method of deep learning (DL), AI methods applied to optimization problems has a longer history, especially when engaging in the famous weapon-target assignment problem (WTA). The more general acknowledgements of the necessity for using AI/ML techniques to optimize targeting in this respect can be found in Boury-Brisset and Berger (2020) framework. It provides a generic solution for “sensor/weapon – target

matchmaking which can accommodate any resource type (kinetic and nonkinetic), as well as multi-objective optimization using genetic algorithms, together with simulation for near-optimal sensor/effector planning, tasking and scheduling” (Boury-Brisset and Berger 2020, 3). More focused research is available, systematically organized and grouped based on similarities and relevance. It offers a concise, non-hierarchical overview of recent WTA studies, highlighting their respective perspectives. Most of these researchers are within the field of operations research (OR). The review was done during the conceptual modelling phase of the experiment and represents the most significant work conducted by other researchers engaged in similar investigations as this experiment. The review is dense and uses terms relating to the field of optimization (Lundgren, Rönnqvist, and Värbrand 2010), but nevertheless provides an opportunity to define the research gap. A good starting point is Quttineh et

al. study on aircraft mission planning to attack five ground targets, where the objective is to maximize the outcome of the entire attack, while also minimizing the mission timespan. (2013, 109) Their use of maximizing one variable while minimizing another is a standard optimization objective (see for instance 33rd Conference proceedings on OR (EURO 2024)). Quttineh et al. (2013) include a synchronization function to differentiate between aircraft used for attacking and those used for illumination of the ground target. The model also includes precedence constraints between targets. A linear mixed-integer programming model is proposed to define the optimal route of aircraft. Interestingly, it takes hours for the solver (CPLEX) to verify optimality for problem instances with only five targets. In another study (Quttineh and Larsson 2014), they 44 | EXPLORING DECISION ADVANT AGES investigate the same problem but applies “metaheuristic solution methods” to solve the problem (2014,

1624). Other researchers such as Ahuja et al (2007) investigate a static Weapon-Target Assignment (WTA) problem in which the inputs to the problem are fixed, meaning all the targets and weapons are known. It is a single stage solution to each assignment, and where the objective is to minimize the total expected survival value of all targets. The problem or decision variable is to determine the number of weapons (i)to be assigned to target (j) subject to the probability (p ij) of destroying target (j) by a single weapon of type (i). They propose both exact (linear and multi-integer algorithms) and heuristic algorithms to solve large-scale instances of WTA in real-time. Andersen et al (2022) also study a static WTA problem similar to Ahuja et al. (2007) in which they convert the objective to a linear function (by developing a compact piecewise linear convex underapproximation). Others such as Lu and Chen in 2021 have studied static WTA problems using integer linear programming instead of

non-linear and present an exact algorithm that utilizes binary columns, and solves the model by column enumeration and branch-and-bound techniques (2021). Ma et. al (2021) study the classical WTA problem but introduce the uncertainty aspects of weapon adequacy that might be caused by dynamic changes in the range of kill probabilities. The problem is formulated as a nonlinear integer programming model with the objective of maximizing the total value of ‘destroyed’. A two-stage hybrid heuristic search algorithm is proposed. Chang etal (2023a) investigate multi-stage WTA problem, which is to assign limited weapons to all targets in multiple attacking phases. The problem is formulated as a nonlinear integer programming model to minimize the total cost of weapons used in all stages. The main decision variable here is the number of weapons (i) that is assigned to target (j) in stage (t). Hocaoğlu (2019) addresses the allocation of surface-to-air missiles (SAMs) to incoming air targets

where the objective is to maximize the air defense effectiveness of land-based air defense systems. Hocaoğlu applies a non-linear, multi objective optimization model, that is expanded to goal programming and accounts for the time dimension by considering engagement durations and setup times between consecutive engagements before deciding the timing of a missile to be fired to the target. In a more recent study, Wang et. al (2023) investigate a multi-objective weapontarget assignment problem for unmanned ground vehicles (UGVs) A non-linear optimization model with objectives of cumulative system revenue and average time cost is being utilized. A new multi-weapon target assignment architecture and a multi-objective artificial bee colony algorithm (MOABC) are proposed to solve the problem. Lai and Wu (2019) propose an improved simplified swarm optimization for a multi-stage weapon-target assignment problem, that also considers cost of weapons. Probability that weapon (w) destroys target

(e) at stage (h) is pre-defined EXPLORING DECISION ADVANT AGES | 45 and independent of stage decisions. Silav et al (2022) investigate multi-objective multi-time period weapon-target assignment problem addressing air targets from the defensive perspective. They apply a typical non-linear WTA formulation, and probabilities remain the same across time-horizon independent of actions in earlier time-periods. They have two conflicting objectives: to maximize the probability of complete success of defenses while minimizing the number of changes in the shoot order of the weapon systems. They propose a new solution procedure to generate updated assignment plans by maximizing defense efficiency while maximizing stability through swapping weapon engagement orders. Part 2 Mapping the targeting trajectory The evolution of targeting is closely intertwined with the evolution of air power. It is also affected by air power strategies. Despite the development of other long-range weapons, air power

has been and to a considerable extent still is, the most critical component of targeting. As suggested by Frans Osinga and Mark Roorda (2015), targeting evolved through the introduction of a third dimension of warfare – air power, giving military organizations the capabilities to target adversaries beyond the frontline of a war. The ability to quickly reach deep into an adversary’s geospatial realm gave birth to new military strategies. Giulio Douhet, an Italian general, was one of the earliest proponents of strategic bombing and targeting of an enemy's industrial and economic centers to achieve victory. Douhet believed that targeting an enemy's vital centers (“Take the center of a large city”) with air power could lead to their collapse and make further ground operations unnecessary (Pape 2014, 60). Even if Great Britain was the first air force to put Douhet’s model into practice in its doctrine, the US also developed concepts for strategic bombing during World War

I as illustrated in Maj. Edgar S Gorrell's 1918 plan “aimed to "wreck" critical targets through bombing to cut off supplies to enemy armies.” (Glock 2012, 147) Developed in the aftermath of World War I, Douhet’s 1921 treatise on strategic bombing ’The Command of the Air’, argued that government, industry and even civilian populations could be considered strategic targets (Douhet 2019), – a concept later seen in practice during World War II and the Korean War (Pape 2014). Other theorists such as Mitchell, an American general and influential proponent for air power in the 1920s and 30s also favored attacking the adversary’s critical capabilities and will to war (Segrè 1992). The strategy of Douhet also highlighted the need for systematic target analysis and intelligence to identify vital targets. Mitchell noted this and highlighted the need for specifically trained target intelligence officers at the staff and unit level, “to compile and maintain all

information of value in the preparation of 46 | EXPLORING DECISION ADVANTAGES bombing missions, an indexed file of photographs, and a stock of maps and charts showing bombing targets and intelligence concerning them.” (Glock 2012, 148) World War II validated the views of early air power theorists on the importance of strategic bombing and targeting. Theories in favor of strategic bombing were, according to some experts (T. D Biddle 2019; Brent 2005) a direct response to the experience from WW I, advocating air power as the means to avoid future wars of attrition (Douhet 2019). The pioneers of air doctrine argued that air power was “more efficient (and therefore more worthy of resources) than sea and land power.” (Pape 2014, 65). Moreover, the air campaigns against Germany and Japan during WWII had created incentives to establish organizations dedicated to collecting target intelligence material and had also shown the importance of analyzing targets. In the decades that

followed air power, and its organization grew in number and means (nuclear weapons), though the previous lessons learned on the importance of unity of command and coordination. The US Air Force made targeting a formal career field and was designated the executive agent for advanced targeting by the Department of Defense, reflecting not only that targeting required specialized training, but more importantly, that the ‘Air Service’ took on the lead in air targeting (Glock 2012). Nevertheless, when the Korean War (1950-53) began, the US “did not possess the organization, intelligence personnel, data base, or target materials needed to support the application of aerospace forces on the Korean peninsula.”(Glock 2012, 157) A similar description of the US insufficiencies are put forward by Osinga and Roorda (2015). Whilst Pape suggests the significant results of air interdiction during the Korean war carried out by the US Air Force, Navy and Marines 10 , destroying the fighting

machine of the adversary enemy (2014, 148–49) Glock paints another picture by referencing two separate studies that evaluated the effectiveness of the Air Force in Korea. According to Glock, these studies reveal that they “only had 53 target folders”, which were out of date and there was an acute shortage of intelligence personnel to support the targeting function (2012, 158). This indicate that the air interdiction strategy may well have been an outcome and approach from the lack of a proper target system analysis and target material production. It also reflects back to Biddle’s notion on how readiness and training affects “sound force employment” (2010, 4). The war in Korea undoubtedly reinforced the need to have a targeting program ready before war commences, which a decade later was further highlighted during the Vietnam war (Glock 2012). The proper selection of vital targets and the target intelligence material production is essential to any successful application of

targeting. It necessitates an organization with trained and experienced personnel in 10 According to Pape, The total number of combat sorties during the Korean war was nearly 740,000, of which half of them were “air interdictions attacking the railways, roadways and predominately just vehicles”. See Robert Pape’s Bombing to Win: Air Power and Coercion in War, Cornell University Press, 2014, 148-149. EXPLORING DECISION ADVANT AGES | 47 both operations and intelligence. The aftermath and the lessons identified from the war in Korea, made the US Air Force create a targets officer career field and expand the scope of a specific targeting data base of potential targets to ensure the adequacy of air targeting materials (Glock 2012, 159–60). However, despite the insights of intelligence’s crucial role in air power targeting gained in the two World Wars and in the Korean war, it was lost in the 1960s, during a Department of Defense restructuring process which included a

centralization of the intelligence function, and the establishment of the Defense Intelligence Agency (DIA). Although the US failed in its targeting strategies 11 in Vietnam, they also learned hard lessons in relation to targeting, especially the intelligence support activities to targeting. Specific training and education in targeting became centralized and formalized. As highlighted in the following citation from the US Air Force doctrine post-Vietnam: The role of intelligence support in the effective employment of tactical air forces is of critical importance. Targeting is the key function and includes exploitation of all intelligence sources for target development, material production, target analysis, recommendations for strike and strike assessment. (Glock 2012, 163) Just as targeting has co-evolved with the advancement of air power, so has the function of intelligence evolved along the practical lessons learned from wartime targeting. The US Air Force and the US DoD agencies

with functional responsibilities have been leading the western doctrinal and procedural progression from the 1950s to today's sophisticated joint targeting process. Moreover, the US political influence over targeting displayed itself throughout the Vietnam War which created a situation of political micro-management and a fragmented targeting process resulting in a decrease in effectiveness (Osinga and Roorda 2015, 40–41). The lessons from Vietnam and the failed Iranian hostage rescue mission ‘Eagle Claw,’ in 1980 forced the US military to undergo a joint operations transformation spurred by the Goldwater–Nichols Act of 1986 (US Congress 1986). This reformation restructured the command and operations of the U.S military, thereby facilitating a more robust joint command structure to enhance cooperation and operational effectiveness across the military services. 11 US conducted two major series of bombing campaigns, Rolling Thunder 1965-1968, and Freedom Train (and

Linebacker) in 1972. According to Pape, US used “conventional coercion through bombing campaigns from the air to compel the North Vietnamese” which completely failed. See Robert Pape’s Bombing to Win: Air Power and Coercion in War, Cornell University Press, 2014, 174 48 | EXPLORING DECISION ADVANT AGES Evolving rationale for the modern joint targeting concept Several parallel processes account for the accumulated reasons behind moving from targeting towards a concept of joint targeting. The efforts to establish jointness in targeting had begun with the centralization of the intelligence function in the 1960s but was further complemented by the Goldwater–Nichols Act, along with new technological innovations such as the precision guided munition (PGM). The technological development in the 1970s and 1980s also included a new range of electronic warfare assets, advanced air-ground surveillance (Joint Surveillance Target Attack Radar System (JSTARS), stealth technologies and

new command arrangements (Finlan 2008). Additionally, the aftermath of the war in Vietnam (1955-1975) added an additional layer of reasons into the already complex decision-making process. New norms were incorporated into targeting based on public opinion and legislation. The preceding air power theories and strategies favoring city or area bombings had led to devastating effects, not just because of their intentions, but for the lack of precision and the inabilities to discriminate targets. Gray (2009), a preeminent scholar and a pioneer in strategic theory, argues that strategies which includes bombardment ‘from the sky’, need to consider the “effects of the damage upon the course and outcome of a conflict” (Gray 2009, 30) . As a response and consequence of such fallacies 12 , novel mitigations began to submerge in order to restrict what was perceived as illegitimate means of warfare (Osinga and Roodra 2015). New establishments within the military’s intelligence was

created to develop models of an adversary’s complex system including its interdependent subsystems 13 as well as forming special organizations to support targeting policies and guidance where lawyers and academics were included (Osinga and Roorda 2015). The old approach was replaced with new theories and strategies that mandated precision bombing and a more systematic approach to reduce costs and manage resources. As a consequence of that, commanders needed to balance how air power resources were employed. This was, and still is, a relevant balancing act or in modern air power terms called ‘air apportionment’, and concerns recommendations on how to use the air efforts to meet a joint force commander’s objectives (US JP 3-30, 2019, 73–75). Deciding whether to direct aerial resources to support ground forces, engage in airto-air combat, or focus on distant targets became a critical factor in effective targeting. Having fewer aerial resources than targets or missions requires

prioritization remains a contentious issue in targeting. The role of the air power also shifted during the 1980s. As articulated in the AirLand Battle doctrine FM 100-5 published in 1982, air power’s role was to “support the ground commander on the battlefield” (Olsen 2017, 40). As a counter-reaction to this 12 The most renowned are perhaps the bombings over Hiroshima and Nagasaki. For more information on fallacies see Colin S Gray ‘Understanding Airpower: Bonfire of the Fallacies’, (Alabama: Air Force Research Institute Papers, 2009). 13 The equivalent of the current Target System Analysis (TSA) EXPLORING DECISION ADVANT AGES | 49 ‘misuse’ of the air powers potential, John Warden, one of the most influential American air power theorists at the time, and the architect behind the air campaign in ‘Desert Storm’ 1991, wrote ‘The Air Campaign’ in 1988, expressing new ways of theorizing the use of the US Air Force. His systematic way of connecting ends (political

objectives), ways (strategies to attain those ends), and means (target prosecutions to meet the objectives) “became a useful guide to planning air campaigns at the operational level of war” (Olsen 2017, 40). Warden also introduced terms such as ‘air campaign’ and ‘center of gravity’, which are commonly used today in doctrines and operational planning documents (2017, Ibid.) Another reason for a joint concept lies in the doctrinal incongruity evident between the services. Since all services had their own air assets, they all had different perception on when and how their assets should integrate in a coherent joint air striking force, and when to support ground forces in combat 14. These discrepancies led to the establishment of joint operations centers (JOC). Specific targeting boards were established where senior staff officers evaluated the situation, assigned priorities and recommended measures to assure the coordinated use of (joint) weapons (Osinga and Roorda 2015).

These changes were thought to mitigate unnecessary misuse of resources In addition, the joint concept of targeting aimed to increase synchronization and to expand synergies from integrations, but also to ensure sufficient deconfliction and coordination between forces during execution. The four terms introduced are important to separate. Synchronization is the arrangement of military actions in time, space, and purpose to produce maximum relative combat power coordination is not concerned with time therefore less complicated, deconfliction separates actions in space to circumvent the need for coordination (Chairman of the Joint Chiefs of Staff 2021, 207), and lastly integration which is, pending the context, either reflecting self-arrangements of forces or reflecting visions of synergies from domain linkages such as UK MOD ‘MDI’ (DCDC 2020). When Operation Desert Storm (1991) began its targeting phase in January 1991, the synergy between operational planning theories and

technological innovations became immediately apparent. (Osinga and Roorda 2015, 48–49) They compare the effectiveness of the ‘new standard, in targeting capabilities developed through Operation Desert Storm with the capabilities in World War II “during 1943 Allied bombers attacked 123 target complexes in Germany. In contrast, during the first twenty-four hours of Desert Storm, 148 target complexities were attacked” (Osinga and Roorda 2015, 42–44). An example of the intellectual tools used at the time is Warden’s five-ring model that enabled visualization of categories of targets, their relative importance and their vulnerability to attack (Olsen 2017). These ideas were further developed by air strategist, and former fighter pilot David Deptula. He offered new perspectives on air power’s unique characteristics such as “speed, range, flexibility, precision and lethality” (2017, 43). Deptula also introduced a viable 14 In contemporary terms this is now referred to as

Close Air Support (CAS). 50 | EXPLORING DECISION ADVANTAGES alternative to attrition warfare and maximum destruction by exploiting the emerging technologies (Deptula 2001, 17). Deptula’s strategy, termed effect-based operations, aimed at achieving the desired effects with minimum damage to the adversary’s infrastructure and civilian population (Olsen 2023, Intro.) Furthermore, he came to “transform intelligence, surveillance, and reconnaissance as well as drone operations” (Ibid. 2023, Intro) Desert Storm demonstrated that the US had institutionalized jointness into its targeting apparatus and it had integrated modern technologies into tangible military capabilities and succeeded in its force employment. Yet, the joint concept of targeting had its flaws and new challenges were on the rise. Contemporary targeting is a complex, methodical cycle and still tightly integrated with the air tasking process. It involves intelligence analysis, weaponeering, force allocation, and

assessing effects: all focused on achieving the commander's objectives through effective targeting. Targeting allows the concentration of a joint force’s versatile capabilities against critical targets to create desired effects. Three factors contribute to the complexities and challenges. First, its role within the larger context of joint operations is to serve as an extension and a command and control (C2) instrument of a joint force commander (JFC) at the operational level of command, directing the targeting efforts performed by the services within a joint force. Second, the challenges of integrating more domains and advanced technologies and third, the complexity of contemporary decision-making within targeting to include human cognitive limitations to apprehend and make timely use of all available data and information in an increasingly digitalized battlespace. Joint targeting as an extension of command and control The task of performing joint targeting implicitly means to

command and control its employment in an operation. From a US doctrinal perspective, joint targeting resides within the joint fires function. Joint functions are elements that founds combat power and include intelligence, movement and maneuver, fires, information, protection, sustainment, and C2. All joint functions are concerned with managing forces / assets in time and space and involves decision-making as well as decisionsupport. The US military doctrine on joint operations from 2018 (US DoD Joint Publication 2018a) defines C2 as: The exercise of authority and direction by a commander over assigned and attached forces to accomplish the mission[.] Control is inherent in command To control is to manage and direct forces and functions[.] Control provides the means for commanders to maintain freedom of action, delegate authority, direct operations from any location, and integrate and synchronize actions throughout the operational area (OA). (2018a, 15). EXPLORING DECISION ADVANT AGES |

51 The importance of clear intent, objectives and explicit guidance has evolved from lessons learned over the last hundred years. Theories on command and control have, however, been developed throughout the history of warfare. Hence, to support this research, the review on C2 and its relations to targeting begins in the 1980s, when sub-tactical technological innovations started creating challenges at tactical and operational levels of command. However, a minor deviation to this will initiate the discussion in order to explain mission command. Limited information and means of communication stimulated the creation of ‘auftragstaktik’ (or mission command). This originated from German tactics in the 1940s supporting commanders to direct forces across a two-dimensional area of operations within mainly a single domain (Jordan et al. 2016) It is still used to encourage flexibility and agility of subordinate commanders’ mission execution as opposed to detailed orders. To some extent

mission command is a semi-autonomous conduct of missions by subordinate commanders within their commander’s intent and guidance of a mission. The concept of AirLand Battle doctrine in the 1980s was an attempt to utilize mission command, but for two domains. It promoted the use of air forces in support of ground forces and was intended to leverage on recent innovations in sensor technology (intelligence, surveillance, and reconnaissance (ISR)) and the new means of communication to facilitate sensor-data on emergent targets reaching decisionmakers in time to ’strike deep’ (Hamm 1988; Jordan et al. 2016) It brought two domains closer together creating a three-dimensional volume of battle space. However, it necessitated more advanced systems to support the proper employment of C2. Hence, the role of technology began to impact how command was executed imposing new frameworks to analyze these challenges. One of the more controversial and influential thinkers in the 1980s was John

Boyd, a fighter pilot and strategist who suggested a new perception on C2 in what he termed ‘Organic Design for Command and Control’ (Boyd 2018). The biological metaphor used in the title was to stress the importance of the human dimension, thereby counterweighting the emphasis that was on technology. His starting point was that, since no foreknowledge, plans, or estimates are perfect, any operation undertaken must allow for adjustments to both tactics and strategy. Hence, Boyd suggested altering command with leadership and control with a monitoring ability, though emphasizing a bottom-up approach (Boyd 2018, 249). Three factors were vital in Boyd’s concept: trust, understanding the commander’s intent and communications to enable initiative, variety and rapidity (Osinga 2007, 199–200). Other researchers on command and control (Mccann and Pigeau 1999) suggested that it should reflect a common intent to achieve coordinated action, thereby defining command as “The creative

expression of human will necessary to accomplish a mission” (1999, 5). Their view of command as being a unique human activity was echoed a decade later by Brehmer (2010), who argued that the whole C2 enterprise starts from the observation that C2 is a human activity that aims at solving military 52 | EXPLORING DECISION ADVANTAGES problems (2010, 2). Brehmer’s point of departure is that C2 is “concerned with design and execution of courses of action to achieve (military) goals” and that it should be perceived as having less focus on people and more on the C2 system as such (2010, 3). Vassiliou, Alberts and Agre (2015) research on network-enabled enterprises, have suggested a similar definition of C2. The prevailing paradigm within western doctrines is related to a human-centric approach to C2 (US DoD Joint Publication 2018a). The intent of the human commander is vital for determining what actions to take and how to propagate that intent. There is, however, a flaw For the

intent to stay valid for more than a few hours into an operation it must be broad, which limits responses to the dynamics of warfare, including how forces and assets are utilized in time and space. Besides, how to propagate the intent implies that communications are unimpeded to access updated information about own forces and assets as well as the adversary. The perception and definition of C2 by Boyd, Brehmer and Vassiliou, Alberts and Agre, is less human-centric. Instead, they focus on the function of command – not who is in command. A new targeting philosophy emerged alongside doctrinal improvements and a refined theoretical understanding of command structures, all aimed at enhancing joint operations. The new strategy combined with the reformation in general in the years before Desert Storm in 1991 had novel applications that transformed the targeting-related command arrangement (Osinga and Roorda 2015, 45–46). By appointing a single commander, a Joint Force Air Component

Commander (JFACC) now controlled all air assets in theatre using Air Tasking Orders (ATO) to enable deconfliction, coordination and synchronization of all air assets. (Ibid 2015) ATOs are still in use today though changes are required based on the requirements of faster processes as stated in a recent US report (Hoadley and Lucas 2018). An air tasking order is defined as a "method used to task and disseminate to components, subordinate units, and command and control agencies projected sorties, capabilities and/or forces to targets and specific missions” (Chairman of the Joint Chiefs of Staff 2021, 20). As such, ATOs are daily orders that articulates airpower strategy and planning efforts for the upcoming 24-hours including support to joint targeting. An additional modernization of C2 was the creation of the Joint Targeting Coordination Board (JTCB), which still is a Joint Force Commander’s mechanism for the control of targeting related decisions. Yet, despite the

technological advancements and doctrinal enhancements, there were internal challenges and room for improvements. Intelligence support for targeting struggled to keep pace with the air campaign tempo, leading to inaccurate reporting, nomination of targets already attacked as well as inconsistent assessments (Osinga and Roorda 2015, 49). As evident from a report after Desert Storm, subordinate commanders had used the opportunity to favor their own service doctrine, rather than the joint doctrine (since EXPLORING DECISION ADVANT AGES | 53 their respective doctrines were contentious at that time) as elucidated in the following citation: General Schwarzkopf believed he resolved the issue by appointing General Homer his JFACC, giving him the authority and complete responsibility to plan and execute the air phase of the campaign. But this initiative did not completely resolve the issue. The Marines retained control of the majority of their air and provided minimal air support to the

JFACC, ensuring this issue did not become significant (Keaney and Cohen 1993, 156). The reasons for these weaknesses were rooted in the doctrines. The doctrines of the individual Services did not align with the intentionally overarching joint doctrines. Additionally, these doctrines had neither been widely implemented nor tested in actual warfare, a fact that became evident during the campaign. On the other hand, this doctrinal muddle of command arrangements for targeting has also been evident in more recent coalition operations. As uncovered during NATO’s engagements in the Balkans and its Operation Allied Force (March-June 1999), it revealed insufficiencies in software tools to plan and communicate, slow processes to confirm target engagement results from Battle Damage Assessments (BDA) and the inability to execute target development (TD) in order to produce and share a consolidated Joint Integrated Prioritized Target List (JIPTL) (Osinga and Roorda 2015, 53–55). Moreover, novel

processes were set in motion in the late 1990s and early 2000s to deal with more dynamic targets emerging within an already decided Air Tasking Order (ATO). These emerging targets required a more flexible method (Osinga and Roorda 2015, 55). It later became defined as the dynamic method in contrast to the pre-planned and cyclic deliberate method (NSO 2021). The twenty-first century wars have also amplified the debate on centralized versus decentralized command and control in targeting. The pace of warfare has accelerated inflicting faster decision-making based on an ever-growing amount of data, information, and intelligence. An important factor in handling emerging potential targets is the push toward decentralization, particularly as expanding targeting domains and dimensions create a broader space for multiple simultaneous threats to appear. NATO’s war in Kosovo necessitated centralized control over flexible targeting engagements and targets may even be perceived as critical for

the outcome of an operation as was the conduct and synchronization of time-sensitive targeting (TST) during Operation Enduring Freedom to mitigate the lack of integration and inefficiency (Osinga and Roorda 2015, 61). These lessons learned on C2 arrangements and centralized tasking philosophy have motivated the control reflected in, for example, NATO’s doctrine today (NSO 2021). 54 | EXPLORING DECISION ADVANTAGES However, the tempo and speed of contemporary warfare has put the current perceptions of command and control under pressure. Neither a single JFC nor a board of senior officers are likely to be able to direct the targeting efforts performed by the services within a joint force in a way that makes the joint force’s targeting efficient and advantageous relative to a peer adversary. Becoming more versatile, rapid and seizing opportunities as they arise requires a targeting process which involves the abilities to understand, decide and take action in a timely manner (B.

Johnson et al. 2023; Brose 2020) This mirrors Boyd’s (2018) idea and core arguments to his OODA loop (Observe, Orient, Decide, Act) for time-sensitive decision-making and his focus on arrangements that enables pertinent command, control and cooperation in complex dynamic situations(Osinga, 2007, pp. 189–190) It also echoes Virilio’s preceding insights of an accelerated culture (2002) and Der Derian’s statement “speed as the essence of modern warfare” (1990, 298). The integration of more domains and advanced technologies The increased integration of domains, forces and assets in the early twenty-first century challenges the abilities to employ joint targeting and to exercise command and control whilst doing it. The recognition and appreciation of additional domains (such as space and cyber) beyond the traditional domains of land sea and airdomains has been formulated as multi-domain integration (MDI). UK Ministry of Defense (MOD) MDI-concept (DCDC 2020) is defined as being

beyond ‘joint’ as this statement reveals; “Our response to the threats, challenges, and opportunities we face is to pursue integration; joint is no longer enough. MDI is more than being good at joint or simply adding space, and cyber and electromagnetic considerations. MDI is about designing and configuring the Whole Force for dynamic and continuous integration of all global capabilities together, inside and outside the theatre, munitions and non-munitions, above and below the threshold of armed conflict.” (DCDC 2020, 23). The US Department of Defense (DOD) ‘joint all-domain operations concept’ (JADC2) is another example of similar strategic efforts for amplified integrations of domains. It is purported to provide commanders with access to information and intelligence enabling simultaneous and sequential operations using surprise, and continuous integration of capabilities across all domains (Hoehn 2021, R46725:5–6). JADC2 is intended to be a “cloud-like environment

for the joint force to share intelligence, surveillance, and reconnaissance data, transmitting across many communications networks” to enable faster decision-making (Hoehn 2021, R46725:2). According to Hoehn, the US DOD has used the ride-sharing service Uber as an analogy to describe EXPLORING DECISION ADVANT AGES | 55 its desired end state for JADC2, and where “the logic would find the optimal platform to attack a given target, or the unit best able to address an emerging threat.”(Hoehn 2021, R46725:1). Not surprisingly, DOD is pursuing three new or emerging technologies to enable this: automation and artificial intelligence, cloud environments, and new communications methods (Hoehn 2021). The current state of the US JADC2-concept described in a summary report indicates an apparent understanding of the challenges ahead, to include information sharing designed and scaled at the enterprise level, security, resilience and cross-domain capability options (Department of Defense

2022, 4). Notwithstanding these efforts to connect forces and assets or to enable fusion of data and information from different domains in real-time, Robert M. Clark (2020), describes some of the challenges as “we cannot simply provide intelligence to customers; they already have more information than they can process, and information overload encourages intelligence failures”(2020, preface). With the movement from traditional military hierarchies to warfighting networks for integration of information across domains, Clark emphasizes more target-centric collaborations to produce actionable intelligence, relevant to the customer needs in conducting operations (2020). The US DOD contends that using one or even two dimensions to attack an adversary is insufficient and that challenging an adversary’s targeting calculus thus requires more complex formations (Hoehn 2021, R46725:5). The increasing complexity, combined with potentially decreasing times to respond to threats from

emerging technologies are likely to require new methods to manage networked forces and increasingly more capable assets. This requires a targeting system capable of facilitating data fusion, combined with sense-making (understanding what needs to be done) and mechanisms for deconfliction, coordination, and synchronization of increasingly digitized forces and assets (understanding how it should be done) to generate effective orders. (Schubert et al 2019). Moreover, speed, not least in prioritization of targets is more relevant than ever. Being able to perform the targeting cycle in a dynamic setting faster than the adversary is not just fundamental to warfare in general, it is essential to the targeting practice. Future conflicts may require leaders to make decisions within hours, minutes or potentially seconds compared with the current multiday process for analyzing the operating environment and issuing commands (Hoehn 2021, R46725:1). These increased demands for integration and the

subsequent development of advanced technologies to enable envisioned warfighting capabilities are challenging, not only the military organizations in their practice, but the very role of the human as being in command and control within the targeting process. The amplified integration of new domains (space and cyberspace), also instigates new means and methods of warfare and widens the scope of war (Yan 2020). In turn, joint targeting may become an epitome of a more comprehensive approach to warfare, integrating 56 | EXPLORING DECISION ADVANT AGES and making use of a vast number of means (lethal and non-lethal) to generate effects on a wider range of targets (to include cyber targets (websites) and space (communication and ISR-satellites). This may create what Penney defines as arrangements of targeting that resemble a redundant ‘kill web’ (Penney, 2023, p. 2) The current targeting challenges necessitate the integration of artificial intelligence to enhance and streamline

critical processes and decision-making loops, thereby creating decisive advantages. This integration will impact the concept as a whole, including its core command and control functions, with significant implications for command structures. How AI became an integral part of military innovation From being more than a periphery portion of computer science for half a century, AI is now perceived as a critical part in most nation’s strategies, and has become a cornerstone in the evolution of warfare according to strategic military policy papers (US DoD 2018; Tarraf et al. 2019) The rationale is supported by academic works (Gill 2019; J. Johnson 2019a; Ernst 2020) and relates to recent advances in machine learning (Schubert et al. 2019) Machine learning allows AI to sense, perceive and reason which, in turn, is opening new areas of potential military applications including symbiotic AI (where AI becomes more of a partner than a tool), deep learning (using neural networks) and

probabilistic computing (NATO Science & Technology Organization 2023). According to Johnson, the current AI surge is due to the convergence of four enabling developments: (1) computing processing power; (2) expanded data sets; (3) advances in machine learning techniques and algorithms and (4) the rapid expansion of commercial interest and investment in dual-use artificial intelligence (2019c, 444). However, AI was initially a separate field of research with few relations to military research and development. The origin of artificial intelligence is, not surprisingly, first found in human fiction. In the 1940’s the ‘three Laws of Robotics’ was introduced by scholar Isaac Asimov in his science-fiction book to reflect some ethical positions with regard to future developments of robots (Asimov 2013). Similar ethical considerations to what Asimov highlighted remain central to discussions some 80 years later. During the 1950s science began engaging in AI from a mathematical and

logical perspective. Although the roots of artificial intelligence can be linked even further back in history, the idea of machines being able to think, were established by a few but important pioneers in computer science, including Alan Turing, John McCarthy and Marvin Minsky (Russell 2019). AI has traditionally been rooted in the field of computer science, but disciplines such as cognitive psychology have significantly contributed to its growth and development (Langley 2012). Defining AI remains challenging, as its meaning has evolved over time, adapting to advancements and shifting perspectives throughout its developmental cycles. EXPLORING DECISION ADVANT AGES | 57 AI has progressed in four significant waves since its beginning. The first wave, known as symbolic AI, emerged in the 1950s and 1960s, and focused on rule-based systems and symbolic reasoning where AI systems were designed to manipulate symbols and strings to simulate human problem-solving (Russell and Norvig 2014,

547–49). It was then perceived as the solution for all logical problems Some of the earlier work on what is today defined as human-machine integration was conducted by US Air Force already back in the 1960s approaching conceptions on how AI could augment human intellect (Engelbart 1962). Minsky explored how intelligence could emerge from non-intelligence, through the portrayal of agents working within a ‘society of mind’ (1986). Despite early successes, the limitations of rule-based systems led to a decline in interest and investments. The second wave, starting in the 1980s, was characterized by the development of expert systems. Two of the pioneers during the second wave were Feigenbaum and Buchanan, who both demonstrated an early potential of expert systems in realworld applications (Buchanan and Smith 1987). The underlying principle of expert systems was that all AI systems represent and use knowledge. According to Buchanan and Smith (1987), the conceptual paradigm of problem

solving that lies behind all of Al is that “a program, or person, can solve a problem by searching among alternative solutions” (1987, 8–9). The expert systems therefore used knowledge bases and inference engines to mimic the decision-making abilities of human experts in specific domains (Gevarter 1982, 6–8). However, to search a defined solution space efficiently and accurately became difficult, largely because of the limitations in computational power. As pointed out in a summary by Buchanan, the number of possible solutions “may be astronomical” pending the number of parameters (1987, 9). The third wave, beginning in the late 1990s and early 2000s, saw the rise of machine learning (ML). This approach shifted focus from explicit programming to datadriven learning where algorithms learnt patterns from large datasets Today, machine learning is an integral, though often unnoticed part of society; from web searches and content filtering on social networks to recommendations

appearing on smartphones. The applications of machine learning (ML) techniques are pervasive, limited only by human imagination and understanding. Increasingly, these applications make use of a subclass of techniques called deep learning (DL). These involve neural networks with many layers and has led to progressions in computer vision, natural language processing and other fields (Lecun, Bengio, and Hinton 2015, 436). As argued by LeCun, Bengio and Hinton, conventional machine learning has been limited in their ability to process raw data, but the emergence of deep learning (DL) techniques can transform raw data into an internal representation that “allows a machine to be fed with raw data and to automatically discover the representations needed for detection or classification.”(2015, 436) 58 | EXPLORING DECISION ADVANTAGES Future trajectories A new generation of weapons systems are being equipped with artificial intelligence, enabling more automation and autonomy. In a

recent news article, US Admiral (ret) Stavridis and former U.S Marine Corps officer Ackerman (2024) suggests swarms of drones will change the balance of military power on the battlefields. They argue that the threat from drones is manageable today, but “when hundreds of them can be harnessed to AI technology, they will become a tool of conquest.”(2024, 1) At some point, IR scholar Christopher Coker argues, drones may take ethical decisions on humans’ behalf, and “programmed with moral algorithms equivalent to the moral heuristics programmed into us by natural selection, there will be little need of pilots” (Coker 2015, 23–24). Even if this reflects a push towards increased autonomy at a lower system level, it still imposes considerations. Jensen, Whyte and Cuomo (2019) argues that AI-enabled weapons are becoming less physically destructive to humans, perceiving that they will be aimed for targeting technical systems in the space domain or transformed into cognitive weapons

using information as munitions by converting it into algorithms seeking targets within the cyber domain. However, contemporary conflicts and wars are suggesting the opposite (Davies, Mckernan, and Sabbagh 2023). Other types of AI-enabled weaponization include recent research on bioweapons targeting specific DNA profiles and genetic setups (E. B Kania 2020) The US National Security Commission (2021) states that the current effort on artificial intelligence is focused towards building software that can do things better than humans or even do things that seem humanly impossible. Martin van Creveld, a military historian and theorist, states in a historical analysis over two-thousand years of war that war is not only “permeated by technology, it is governed by it” (van Creveld 1989, 1). This argument is counterbalanced by other, more irrational factors such as honor, loyalty and sacrifice (1989, 314); nevertheless, his research highlights how technological innovations are assimilated

into military concepts and, through organizational adoption lead to changes in how wars are fought. Visions of near future warfare do not seem to challenge Van Creveld’s notion of the power of technological innovations. On the contrary, scholarly work on future military concepts suggests that advanced technologies spearheaded by artificial intelligence (AI) will spur more automation and autonomy that may have a revolutionary impact on warfare, modifying the character or even the nature of war (B. M Jensen, Whyte, and Cuomo 2019) AI’s transformative potential challenges longstanding and foundational warfighting principles. Experts in the field such as Ayoub and Payne, Johnson and Del Monte agree that AItechnologies are increasingly intelligent and that the digitalization of societies and subsequently military organizations enhances an interconnected, distributed and digital near future of warfare (Ayoub and Payne 2016; J. Johnson 2019a; Del Monte 2018). EXPLORING DECISION ADVANT

AGES | 59 The strategic efforts to integrate the services (Army, Navy, Air Force) and the domains (land, sea, air, space, and cyberspace) in which warfare is executed are largely visions today. These aspirations appear to seek advantages from the growing digitalization of the world. The digital age provides access to borderless information and abilities to share information that transgresses traditional boundaries (army, navy, and air force) in warfare. Furthermore, technological innovations are often portrayed as offering opportunities to those that embrace them, as captured in some of the US Defence AI-related programs and innovation initiatives, to include the Third Offset Strategy, Project Maven, DARPA’s ‘AI Next Campaign’, the establishment of the Joint Artificial Intelligence Center (JAIC), the Joint Common Foundation JCF, and the DoD’s ‘AI Strategy’ (J. Johnson 2019b, 10) These ideas focus on securing future operational advantages by integrating innovative

technologies into new operational concepts that address key challenges. Ambitions to offset an adversary is certainly as old as warfare itself, but crafting superiority in a future digitized complex battlespace is not an easy task to undertake. Mcgrath, outlines a number of information operations (IO) considerations, suggesting that the problem set of future warfare’s multi-domain integration is far more complicated than many would like to believe (2016, 22). To develop systems that can facilitate adaptive warfare in which a commander’s intent and guidance can be operationalized through different capabilities by different commanders at all levels against an adversary’s weaknesses appears hard without investing in both technologies and a new mindset. Conducting highly integrated joint operations that are synchronized with other ongoing joint actions are likely to necessitate delegating authority further down the levels of command, and on in a wider manner. As McGrath suggests,

achieving this will require a research environment that leverages decision-support modelling and simulation (M&S) techniques: these include AI and ML software, agent-based modelling, augmented reality, system dynamics, and game-theoretic modelling within a federated architecture (2016, 20). This is required to accurately model complex and adaptive systems which can facilitate the understanding of the challenges ahead (Mcgrath 2016, 20). Strategic ambitions on high-end technologies have also accelerated research and development in other major competitive states such as China and Russia (E. B Kania 2020; Vinci 2019; Cave and ÓhÉigeartaigh 2018). China’s General Secretary Xi Jinping has declared that “only the innovators win [this intense global military competition]” (E. B Kania 2020, 83) and in response to this imperative, the Chinese People’s Liberation Army (PLA) is seeking to improve its capacity to leverage academic research and commercial interests through a national

strategy of “military-civil fusion” (E. B Kania 2020, 83) Russia’s ambitions are similar, and as stated by President Putin “Artificial intelligence is the future[]Whoever becomes the leader in this sphere will become the ruler of the world.” (Del Monte 2018, 185) Confirmations that Russia and China recognizes the potential in defense applications where artificial intelligence and autonomy are at the fore is also evident in a recent 60 | EXPLORING DECISION ADVANTAGES Swedish report from its Defence Research Agency (FOI) (2020). Other researchers such as Ernst (2020) argue that ambition is one thing and that China’s alleged AI acceleration might stagnate due to China’s heavy reliance on foreign sources for AI technology (2020, 23). The suggestion is that China’s lack of a robust body of domestic research has become a major vulnerability for China’s AI industry (2020, 23). Research and development (R&D) advancements are progressively enhancing the probabilities

of more sophisticated, complex undertakings. Johnson (2019a) points out that China’s ambitions is to integrate AI-applications to support “military planning, operational decision-making, and the establishment of joint operations command system to augment and integrate these capabilities.” (J Johnson 2019c, 153). According to Johnson, this is an indicator that AI-applications are not constrained to weapon systems or robotics but rather expanded to artificial systems that are prophesied to augment human decision-making at levels residing on the joint operational command. China also presents clear ambitions to tie together strategic objectives and tactical weapon systems through artificial systems at the operational command level. Johnson makes convincing arguments for the reasons behind this such as increasing range, accuracy, mass, coordination, intelligence and speed at system level in a future conflict, but also warns of the risks associated with ‘nascent technology’ to

include amplified uncertainties and the introduction of new threats to the security landscape (2019a, 159). Investments in this field can also be found in smaller states that anticipate military advantages as well as reimbursements towards state security (Boulanin and Verbruggen 2017; UK MoD 2018; Gill 2019) as well as in larger bodies such as the European Defence Fund (EDF) where multinational joint defense industrial projects are scheduled to commence in 2021 (FOI 2020, 12–13). It is reasonable to state that not only powerful states, but other nations and at least two international organizations (EU, NATO) recognizes the transformative and military-technical potential that artificial intelligence may offer for national security and strategic rational (J. Johnson 2019a) At its heart resides the idea of handing over authority from humans to machines. Some renowned researchers on AI’s effects on war and warfare (Del Monte 2018; Freedman 2017; Scharre 2018b) have written books on

these matters and predict uniformly that handing over authority to machines will likely mutate warfare to become more automated and distance humans farther away from the battlefields. Cronin, an expert in the field of security and technology, worries that the emerging, and open technologies of AI will intensify global military competition and impose an increased civil-military cooperation that in turn may provide non-state actors access to open technologies, unintentionally arming the terrorists of tomorrow (Cronin 2019). Advanced technologies therefore play an important role in the strategies shaping future military concepts (Finlan 2021). Johnson (2019b) argues that some (South Korea, Singapore, Israel, France, and Australia) could become pioneers in cuttingEXPLORING DECISION ADVANT AGES | 61 edge dual-use AI-related technology and key influencers modelling future security, economics, and global norms. Similarly, NATO’s Strategic Trends report perceives 5G as the foundation

for the concept of ‘sensors everywhere’ and refers to the ability to detect and track any object from a distance by processing data acquired from high tech, low tech, active and passive sensors (NATO 2020). This could facilitate a substantial step towards full autonomy of artificial systems. Moreover, the data transmission speed of 5G will enhance connectivity and could enable close-range military communication to function independently of satellites (J. Johnson 2019b) Future growth of 5G combined with big data and advanced analytics (BDAA) could incorporate an effective Internet-of-things (IoT) (Sarker et al. 2024, 16) where everything that can be connected is connected providing networked military applications and services. Conclusion This chapter sought to bridge the gap between AI and targeting literature by integrating these two perspectives. It first set out to chart the current research on the matter. This convergence stems from human cognitive limitations and efforts to

enhance the effectiveness and efficiency of decision-making processes. The reasons relate to either human cognitive limitations or simply attempts to make the decisionmaking processes more effective and efficient. Military organizations and their research indicate an increased interest in AI realizing that the research, methods, techniques and applications can alleviate current challenges ahead. Central to these advancements are deep learning (DL), novel multi-aspect rule-based AI, and context-aware AI. The shared goal remains consistent: optimizing the best possible solutions to well-defined problems. Several considerations are brought to attention when reviewing the literature on decision support from AI. Explainability and transparency are important aspects, not least in critical and sensitive contexts such as targeting. AI in support of human decision-making also has an important ethical dimension to it. However, the review indicates two major ethical standpoints: one arguing for a

more nuanced approach to clarify what human control is and how accountability can be applied; the other is a position arguing the irresponsibility of not using AI in military decision-making processes. These are not necessarily contradictive, rather they represent two viewpoints with a shared notion that AI is inevitably entering into the body of military decision-making. The intelligent agent can and will be an integral part. It is the application of AI, and the considerations made when building them that is of concern. Speed or accelerated tempo relative to an adversary appears to be a common factor. The ability to make faster decision than the adversary is what is driving this development. It is an 62 | EXPLORING DECISION ADVANTAGES apparent and prevailing challenge to contemporary targeting. The narrow state-ofthe-art research review focused on research closely related to the two experiments The first review on applied deep learning and the use of semantic segmentation

highlighted its current applications and identified a clear research gap: its potential for intelligence collection, particularly in optimizing sensor allocation for valuable assets, remains underexplored. General frameworks were apparent, but no actual applications. The second experiment, which focuses on optimizing weapon-target assignments, is supported by a significantly more extensive body of prior research. Especially the novel use of multi-aspect rule-based AI that Sarker et al. presents in their research (2024). The chapter also included a more complete description of the evolution of targeting from being a separate thought new domain dominated by air power to becoming an integrated part of a larger joint concept. It also elaborated more on how AI became an integral part of military research and development as well as a brief discussion on future trajectories. AI is thought to enhance efficiency in processes and workload in several areas, including targeting. Processing data

and information in support of decision-making in an environment that tends to move faster motivates military organizations to adopt AI. Bias and risk are issues that have to be managed and mitigated along the way. When reviewing the perceptions, ambitions, and future investments in AI for military power, it becomes apparent that AI is a prerequisite for most of these areas of military capability development (Eric Schmidt et al. 2021; NATO Science & Technology Organization 2023). Moreover, the use of unmanned platforms is rapidly changing the context of how military power can be utilized (T. Wang et al. 2023; UK Ministry of Defense 2016) However, it remains to be seen whether technology will have such a profound impact on military power. In an authoritative study of military power, Stephen Biddle (2010) suggests that the relationship between technology and force employment is largely misunderstood and that proponents of a revolution in military affairs (RMA) are mistaken (2010, 4).

Biddle defines military power as ‘capability’, an ability to project and use it for political reason in military strategies (2010, 5). Developing tangible military capabilities with AI as an integral component, according to Biddle’s definition, involves more than simply integrating technology into existing systems or concepts. Successful integrations of AI and the development of real capabilities will require adaptation in many areas including the command-and-control structures, the decision-making, and the way (methods and tools) the capabilities are employed and practiced 15. The review also reveals efforts to look beyond traditional modes of warfare, perceiving advanced technology as something more than just equipment. Technology is “rather an expression of choice” 15 In this thesis practices refer to how specific tasks or functions are executed, including the methods and tools used and the approach to how it is carried out or performed through defined procedures and

processes. EXPLORING DECISION ADVANT AGES | 63 that has marked effects on strategy altering “the ‘collectively accepted’ boundaries of the possible” (Finlan 2008, 97). To summarize, AI has significantly transformed decision support systems, from early expert systems to sophisticated models incorporating machine learning, deep learning and other AI-techniques. The development of advanced AI techniques can enhance the accuracy, efficiency, and reliability of DSS across various domains. As AI continues to evolve, the integration of ethical considerations, explainability, and sustainability will be imperative in developing robust and trustworthy decision support systems. 64 | EXPLORING DECISION ADVANTAGES EXPLORING DECISION ADVANT AGES | 65 66 | EXPLORING DECISION ADVANTAGES Chapter 3 - Theoretical framework “There are normally two reactions to what I have set forth in this article. One is, “We think this way already, but our thought processes are quicker,

simpler and more natural.” To this I say, Really? Show me” Wass de Czege, in (Richards 2012, 24) Introduction The chapter sets out the theoretical framework of this project. It provides the foundation from which the project explores the research question using the methods discussed in Chapter 4. The framework draws on three concepts: Boyd’s strategic theory and OODA loop model; the joint targeting process as defined in military doctrines by NATO and the US Department of Defense; and the intelligent agent as defined by Russel and Norvig (2014). Boyd’s “strategic thought” (Osinga 2007, Intro) is a theory. Theories represent what political science theorist Van Evera argues are “general statements that describe and explain the causes or effects of classes of phenomena” (Van Evera 1997, 7–8). Doctrines, by contrast, represent a “storehouse of analyzed experience and wisdom” to solve military problems (Lemay Center for Doctrine 2020, 4). In other words, they define

what a military organization considers to be true about the “best way” to do things (Lemay Center for Doctrine 2020, 9). Furthermore, doctrines are implemented into a military organization as guidance, and can, therefore be distinguished from theory, especially if the content involves novel thinking that may challenge prevailing assumptions or truths. Such is the case with Boyd’s conceptualization of strategic thought 16. Tensions in military organizations could arise when facing new thinking and ideas as these challenge traditions and the stability of old notions. Military organizations may react in the way Brigadier General (ret.) Wass de Czege describes it in the citation This highly innovative thinker 17 was well aware that the preservation of established knowledge sometimes 16 Boyd challenged contemporary paradigms with his theory, including his ideas on command and control. This can be hard for an organization to fully accept and embrace. See for instance Thomas Kuhn’s

“The Structure of Scientific Revolutions” from 1962 (Chicago: The University of Chicago Press 2020), further exploited in Shapere’s “The Character of Scientific Change”, (Dordrecht: Springer Netherlands 1984). More recent work on attempts to uncover the mechanism of scientific change is suggested by Barseghyan, “ Redrafting the Ontology of Scientific Change,” Scientonomy: Journal for the Science of Science 2018 Vol.2 pp13-38 17 Wass de Czege was the principle author of the 1982 edition of FM 100-5, AirLand Battle, after which he became the first director of the School of Advanced Military Studies at the United States Army Command and General Staff College in 1983, Source: AUSA, “Huba Wass De Czege”, https://www.ausaorg/people/huba-wass-de-czege-1, accessed 2024-11-26 EXPLORING DECISION ADVANT AGES | 67 hinders the adoption of new practices. This may lead to dismissing/downgrading new concepts as minor variations of existing practices. Boyd, whose theory is at the

center of this chapter, encountered similar institutional barriers when presenting his ideas on strategic theory to the U.S military establishment in the 1970s and 80s Decades later, most of his ideas have been adopted to complement other concepts. The project uses Boyd’s theory to situate the experiments within the larger context of military decision-making and its prerequisites. Subsequently, the doctrines are used to define the contemporary approach to targeting in a joint force context as these provide clear guidance towards what joint targeting is and how it is done. The concept of the intelligent agent is used to clarify what the treatment in the two experiments actually represents. The three concepts supplement each other Boyd’s theory is not only descriptive, but also prescriptive in nature enabling the project to expound Boyd’s theoretical rationale and suggestions on how military organizations should conceptualize and apply their own OODA loop frameworks to increase

their capabilities to learn and adapt to changing circumstances. Furthermore, as doctrine informs us how contemporary joint targeting should be conducted, this becomes a baseline condition for the experiments. The contemporary literature (as referenced in Chapter 2) underlines the importance of joint targeting becoming faster and more agile. This is what Boyd states as necessary (see Theory section in this chapter) In other words, the project uses Boyd’s theory to infer the need, the doctrines to frame the real problem, and AI (as the intelligent agent) applications to solve them. Collectively, the three concepts serves the experimental setup where the theory is employed as a recipe to improving human decision-making. It provides for a comprehensive understanding towards “organizational adaptability“– to “observe, learn, and adapt” in order to survive (Osinga 2007, 237). The quest to improve the capability of adaptation is a central theme in Boyd’s theory. A capability to

learn and adapt can in this case bring about novel suggestions on how AI applications can be applied to improve joint targeting. Due to the paucity of scholarly publications by Boyd, secondary literature such as Osinga’s authoritative study (2007) will accompany the original published presentations and compilations by Boyd (Boyd 2020, 1976, 2018). To explain the concept of joint targeting, NATO and US doctrines are analyzed along with the relevant scholarly publications. The chapter begins by defining theory and doctrine It then proceeds to explain the two concepts in detail. The chapter ends with a concluding section reflecting on the framework and discussing alternative theoretical approaches that could have been possible to use. 68 | EXPLORING DECISION ADVANTAGES Defining and using theory and doctrine Before discussing the concepts, the chapter introduces a definition to both theory and doctrine. There exists a general consensus that theory should describe and explain

phenomena (see above, (Van Evera 1997)). Some theories have a narrower scope than others, and there may well exist conflicting theories within any given research field. This could originate from new theoretical propositions challenging existing research paradigms (T. Kuhn 2020; Shapere 1984) which is what moves research forward (Shapiro 2005; J. Law 1992) However, theories and its propositions should be testable. As emphasized by sociologist Swedberg, “theory refers to logically interconnected sets of propositions from which empirical uniformities can be derived” (Swedberg 2014, 4). Barseghyan argues that “[a]ny general theory that attempts to describe and explain changes in a certain domain usually contains two major components – a certain ontology of the entities and relations that undergo change in that domain and a certain dynamics of how these entities and relations supposedly change through time” (2018, 14). Barseghyan’s notions of changes in a certain domain reflects

the work of Boyd, in particular through his “core arguments” advocating changes (Osinga 2007, 184–88). As Osinga put forward, Boyd’s work constitutes a theory “of considerable sophistication, consistency and persuasiveness”(2007, 235), and therefore goes beyond being descriptive and explanatory. Its propositions and core arguments are to some extent prescriptive in nature, informing military organization on how to pursue “rapid decision-makingfor success” and how organizations can learn, “adapt and thus evolve.” (Ibid 2007, 256–57) The second concept of the framework is doctrine. Doctrines express fundamental principles about warfare that a military organization has developed and officially approved as a coherent “body of knowledge” (US DoD ADP 1-01 2019, 10). They represent “institutionalised beliefs about what works in war.” (Høiback 2011, 897) and implicates an authoritative, and approved guidance that nonetheless requires “judgment in

application” (US DoD ADP 1-01 2019, 10). In each specific doctrine, related terms and symbols are defined which establishes a common language and enables clear communication. Doctrines may include tactics, techniques, and procedures (TTPs) which are specific guidance and methods for implementing the doctrine. Doctrines can have different intentions of use As pointed out by Høiback, the utility of doctrine can be “a tool of command, [] tool of education [][or] a tool of change”, but should not be all of these at once (2011, 888). Høiback suggests that the main difference in-between the three concepts is that the first two refers to present being, describing “what to do” (command) or “what we do and why” (education), and the latter informs of “what to be[come]” (2011, 888). The evolving knowledge and collective wisdom from operations can and should influence current doctrine and shape it into a best practice. Importantly, since doctrines are authoritative and

consequently constitute a baseline for what joint targeting is, how it is conducted, and the reason behind it, it serves the purpose of framing a baseline EXPLORING DECISION ADVANT AGES | 69 condition for the experiments. The standardized work of doctrines also provides an explicit standard to test solutions for improvement (Richards 2011). Valid doctrines and their specific instructions and procedures represent professional knowledge towards conduct of joint targeting (NATO 2019; US 2020) Hence, the doctrines constitute guidance to how the concept of joint targeting is intended to be applied. Consistent analysis through a systems concept To enable a consistent analysis of Boyd's OODA loop and the joint targeting process, the project treat each as a system with two levels of abstraction. Each system consists of interacting components that work together to produce an overall effect. Specifically: Boyd’s OODA Loop is a system comprising four components; and the joint

targeting process is a cycle consisting of six phases. Additionally, the explicit dynamic targeting method is also a system composed of seven elements. The intelligence support activities associated with the dynamic targeting method is not analyzed this way, as it is perceived as a subordinate activity to the dynamic method. By employing a systems approach in the analysis, it is possible to distinguish between the system as a whole and the system as a set of interacting components. These two levels represent two lenses through which the investigation can be logically divided. As modelling theorist Ritchey (1991) explains: “it is this distinction between system levels -- between the behavior of the system as a whole and the specific relationships between its parts -- which is fundamental to the concept. The idea of a system would be meaningless without this distinction.” (1991, 6) According to Ritchey, the fundamental question of what the system does or accomplishes at the unit

level needs to be differentiated from what the various component within the system does or accomplishes. This systems approach allows for a transparent analysis of Boyd’s OODA loop; the joint targeting process; and the method of dynamic targeting. Furthermore, the approach supports the subsequent synthesis and operationalization of the three concepts as systems, as explained in Chapter 4, where the three systems are compared illustrating their interrelations to each experiment. Boyd’s military thinking and the OODA loop This section introduces Boyd’s military thinking, describes the OODA-loop, its components, and the key themes of Boyd’s strategic theory. The OODA loop was introduced by Boyd in 1977 and is one of the most influential military ideas of the modern age. It is informed by a comprehensive historical analysis of recurring 70 | EXPLORING DECISION ADVANT AGES patterns in warfighting and still provides an inspiring intellectual fountain for operational art and

tactical thinking. Moreover, the OODA loop has been commended in two biographies (Coram 2002; G. Hammond 2001) and thoroughly explained in the authoritative study of Boyd’s strategic theory by Osinga (2007), a Professor of War Studies. Boyd’s influence has also been praised by the renowned strategist Colin Gray: “Boyd’s loop can apply to the operational, strategic, and political levels of war[] The OODA loop may appear too humble to merit categorization as grand theory, but that is what it is. It has an elegant simplicity, an extensive domain of applicability, and it contains a high quality of insights”(1999, 90–91). Here, Gray states that the utility of the OODA loop may result from the combination of simplicity and high quality insights. However, simplicity can also be a pitfall as it can appeal to practitioners’ preconceptions and potentially lead to misinterpretations of the theory’s origins and insights. The selection of Boyd’s theory can be justified by five

arguments: first, the OODA loop concerns the issue of winning or losing, which is still important in contemporary warfighting 18; second, the ability to interrelate to the joint targeting cycle; third, it support rapid cycles making it apposite for the dynamic targeting method (F2T2E2A); fourth, its relevance to date and contemporary military challenges (Osinga 2007, 252–57); and last, its explanatory power of knowledge production via external and internal interactions. From a fighter pilot’s perspective to a strategic outlook John Boyd started his military career in 1945 as an enlisted soldier in the US Army and transferred to the US Air Force where he served as a fighter pilot. During the Korean War he was assigned to the 51st Fighter Interceptor Wing in the winter of 1952-53, flying twenty-two combat sorties in the F-86 Sabre. This was also the period in which he began his assembly of thoughts on air-to-air tactics (commonly known as dogfighting), aircraft design and the OODA

loop to articulate his thinking of strategy and what later became “his thought on time and thinking itself” (Osinga 2007, 20). In the decade that followed, Boyd’s warfighting experience crystallized his ideas, specifically in terms of insights gained from being a fighter pilot and conducting air-to-air combat displayed in his Aerial Attack Study. He could show “that superior maneuvering capability combined with better training and cockpit design” offered advantages in time and situational awareness (Osinga 2007, 25). From the mid-1960s to the mid-1980s he worked at the Pentagon and was involved 18 See Penney ’ Scale, Scope, Speed & Survivability: Winning the Kill Chain Competition,’ Mitchel Institute 2023. EXPLORING DECISION ADVANT AGES | 71 in the development of advanced aircraft such as the F-18 and A-10. This also created opportunities for him to engage in military history. After his role in the development of the attack aircraft A-10, the F-15 Eagle, and the

F-16 Fighting Falcon, his interest arose in investigating the essentials of victory. These thoughts later emerged in Patterns of Conflict (1986), that were strongly inspired the ancient Chinese strategist Sun Tzu. This also layered a path towards explaining winning and losing wars and the assembly of a grand theory. Influenced by early achievements in air-to-air combat and his explorations on military history, Boyd’s thought transcended into a more general concept of warfighting, thereby applicable to all military operations (Boyd 2018, 24). Boyd - a magpie searching for novel ideas Boyd was a magpie who picked up military concepts, themes, and ideas from a variety of areas, which progressively evolved into his own strategic theory. Importantly, Boyd’s ambitious efforts occurred “at a time when the US military was searching for novel ideas to solve concrete strategic and operational problems.” (Osinga 2007, 51). He engaged in a cross-disciplinary and iterative scientific

approach to understand the reasons behind winning and losing. It appears as if Boyd intuitively realized that to reach novel insights, one is required to first analyze the multi-dimensional facets of a problem through diverse conceptual lenses before insightful and novel knowledge can be synthesized. His thoughts resulted in a compilation of arguments in ‘Deconstruction and Creation’ (Osinga 2007, 138–39). The ideas and reconstructions of thoughts also made its way into refinements of the OODA loop. As an example of this, he came to reconsider time and tempo Instead of perceiving that winning a ‘dog-fight’ is to operate at a quicker tempo, not just faster than an adversary, Boyd realized that tempo was more an expression within the larger context of adaptation (Osinga 2007, 28), which later became one of his key themes (Ibid. 2007, 237) His framework, including the early versions of what later became his famous OODA model, was arguably a rather hefty synthesis of a range of

strategists, theorists, and practitioners of warfare to include Sun Tzu, Clausewitz, T.E Lawrence, JFC Fuller, Julian Corbett, and Liddell Hart (Osinga 2007, 29). A compilation of the extensive range of influencers has been masterfully illustrated by Osinga (2007) in his study of Boyd. For this project only a few are highlighted as their voices are clearly evident in Boyd’s OODA loop (Boyd 2018, 383–85). Lawrence’s ideas, for example, reinforced Boyd’s belief in speed and mobility over hitting power and informed him that strategic theory is always contextual and therefore always changing. Liddell Hart’s writings also paved a way from attrition warfare towards a more indirect 72 | EXPLORING DECISION ADVANT AGES approach 19, but also informed that “adaptability is the law which governs survival in war” (Osinga 2007, 35). Adaptation became a key theme in Boyd’s ideas However, the most important influencer forming Boyd’s conceptions appears to have been the ancient

strategist Sun Tzu and his work The Art of War (Griffith 1963). Osinga (2007) has suggested that Sun Tzu was Boyd’s ‘conceptual father’ (2007, 35). Accordingly, Sun Tzu’s ideas are most relevant when analyzing how the deeper meaning or the outer frame of the OODA loop came together and it is helpful to refer to Griffith’s (1963) well-known work on Sun Tzu in the 1960s. The essence of Sun Tzu’s philosophy of war is evident in Boyd’s OODA loop (Osinga 2007, 36–37) and includes the fundamental estimates of the situation beforehand (Griffith 1963, 63). Even though the original five factors 20, or matters (‘shih’) may have changed over the course of 2,000 years, the appreciation of the effect of external and internal environmental factors is still relevant today. For a correct estimate of ‘shih’ requires foreknowledge: an accurate as well as truthful interpretation of the situation including intelligence. The quality of the estimation is dependent on foreknowledge,

which in turn is based on the ability to discern and properly judge the information retrieved from a holistic view coupled with a deep understanding of its meaning. All of Sun Tzu’s stratagems stem from the assumption that one can shape the enemy by using of multiple methods affecting the mental, moral and physical aspects of the enemy’s system (Griffith 1963, 41–43). However, the ways in which one applies the recommended tactics and methods in a synchronized manner is not merely resulting from what effects one strives to accomplish. It is from the superior estimation of the situation at hand. The foreknowledge that enables for the sound estimation also informs the decisions on methods, risk calculations, and the subsequent actions taken. Conversely, as pointed out by other researchers interpreting Sun Tzu’s concept, applying combinations of orthodox and unorthodox methods whether direct or indirect is likely to be insufficient without knowing how to target the enemy (O’Dowd

and Waldron 1991). In this context, Sun Tzu discusses how the two tactical ‘instruments’ cheng (direct) and ch’i (indirect) are compared to two interlocked rings, with infinite possible permutations (Griffith 1963, 43). The latter (ch’i) is always unexpected, strange, or unorthodox, whilst the former (cheng) is more obvious and thereby orthodox. 19 The indirect approach is described as a method emphasizing ‘paralysis’ through ‘brain warfare’ (aiming at mental and moral objectives) by psychological ‘distraction’ and physical ‘dislocation’ of the enemy. See Osinga ‘Science, Strategy and War: The Strategic Theory of John Boyd’, Routledge, 2007, 34. 20 On page 63 in (Griffith 1963), Griffith describes five factors: moral influence, weather, terrain, command, and doctrine. Moral influence refers to the morality of government, a justly cause, and doctrine has a primarily meaning of law or method. See Griffith ‘Sun Tzu: The Art of War’, Oxford University

Press, 1963, 63. EXPLORING DECISION ADVANT AGES | 73 As posited by Griffith, it is misleading to relate the two terms to tactical battle groupings only because these can be launched on strategic levels as well (1963, 44). Hence, the importance of Tzu’s estimate on strategies and tactics. When various methods are simultaneously applied in ‘novel combinations’ (Sawyer, 1994, p. 147) and reinforced by modes of behavior, surprise is achieved. Moreover, it forces the enemy to react and respond outside of existing plans. They become confused and off-balanced leading towards ‘disharmony and chaos’ (Sawyer 1994, 190–91). Sun Tzu’s theory of adaptability dovetails with Boyd’s concepts. As Griffith explains, adaptation to existing situations is an important aspect of Sun Tzu’s theoretical perspective on strategy: Just as water adapts itself to the conformation of the ground, so in war one must be flexible; he must often adapt his tactics to the enemy situation. This is not

in any sense a passive concept[]. Under certain conditions one yields a city, sacrifices a portion of his force, or gives up ground in order to gain a more valuable objective. Such yielding therefore masks a deeper purpose, and is but another aspect of the intellectual pliancy which distinguishes the expert in war. (1963, 43) This suggests that Sun Tzu perceived flexibility as an essential aspect of adaptation. It emphasizes responsiveness and other vital abilities that are central to both Tzu's concept and Boyd's theory. Additionally, Boyd states that “adaptability implies variety and rapidity,” which in turn enables unpredictability (Boyd 2018, 221). Adaptation is a key theme in Boyd’s analysis Borrowing the perception of Sun Tzu, Boyd suggested that the ability to adapt to the situation is key, so that one can take “full advantage of the defining circumstances”(Osinga 2007, 38). Boyd adopted the concept of adaptation into his own design leading towards the

framework of the OODA loop and it became the key theme in his theory. Gaining insights and inspiration by clustering relevant elements from seemingly disparate fields including quantum mechanics, cybernetics and psychology was part of Boyd’s scientific ‘fountain’ (Osinga 2007, 52). However, other theorists of his era also favored the notion of “seeking to be integrative” (Mintzberg, Ahlstrand, and Lampel 2009, 1998:6) as a vital part of the strategy-making process. In his subsequent work Patterns of Conflict (Boyd 2020) based on military history, Boyd sought to highlight continuity and recurring patterns mainly from those that focused on: 74 | EXPLORING DECISION ADVANT AGES “the mutual processes of adaptation, on perception, on the mental and moral impact of one’s own moves, feints and threats, and on achieving destabilizing effects throughout the enemy system [] display[ing] a balanced understanding of the cognitive dimension, in concert with the physical” (Osinga

2007, 29). Furthermore, Osinga reveals the larger context of adaptation and the complexity of Boyd’s rationale on the concept of adaptation. According to Osinga, it was not just winning a dogfight in the air against an adversary. Neither was it pure tactics The OODA loop incorporated adaptation as a central pillar. Employing a systems approach whilst using the OODA loop meant considering all aspects (moral, physical, and cognitive) to overcome challenges whilst disrupting and destabilizing the adversary in all dimensions possible. Adaptation was an ability that should be present throughout an armed service. From the pilot and soldier to the service as a whole. It was integral to both tactics and strategy This suggests that the OODA loop is more than just the apparent use of rapid cycles for decision-making at a tactical and operational level of warfare. Rather, it should be employed to understand and appreciate the process of adaptation. The process is a representation of the OODA

model and for how to continuously acquire and produce knowledge whilst interacting both externally and internally. The conception itself is arguably an ‘epistemological statement’ (Osinga 2007, 242). Since epistemology is defined as “the study of the nature, origin, and limits of human knowledge” (Al-Ababneh 2020, 78), Boyd’s idea of adaptation indicate a strong emphasis on the importance of combining the scientific methods as a guiding principle for novel insights. Building the system – the OODA loop Boyd built his system of thoughts by integrating three factors: first, his experience as a fighter pilot and being part of a military system undergoing fundamental challenges; second, his extensive study of military history; and third, his engagement to other sciences (Osinga 2007; Boyd 2018). His own military experience led him to a convincing idea of the need for a radical change in contemporary military thinking (Boyd 2018), and his intellectual engagement with ideas and

concepts external to his discipline strengthened his own arguments. The similarities and inspiration from Sun Tzu’s The Art of War, for example, can be seen in the prime role of ‘foreknowledge’, the concept of the “orthodox (cheng) and the unorthodox (ch’i)” methods, the role of ‘perception’ and ‘pattern recognition’, and the significance of ‘tempo and surprise’ to shape and create favorable conditions (Osinga 2007, 39–41). The most complete version of the system (OODA loop, Fig. 1) as a whole was compiled by Boyd in 1995 (Boyd 2018, 384), two decades after he first began EXPLORING DECISION ADVANT AGES | 75 studying the essentials of military victory. Osinga suggests that Boyd created a “synthesis of military history and strategic theories” inspired by authors that were “united in their focus on the mutual processes” of adaptation, on perception, and the importance of the cognitive dimension supplementing the moral and physical aspects “of

defeating the enemy in battle” (2007, 29). Figure 1: Boyd's illustration of the famous OODA loop.(Boyd, 2018, p 384) Without analysis and synthesis across a variety of domains or across a variety of competing independent channels of information, we cannot evolve a new repertoire to deal with unfamiliar phenomena or unforeseen change. Without a many-sided implicit cross-referencing process of projection, empathy, correlation, and rejection (across many different domains or channels of information), we cannot even do analysis and synthesis. Without OODA Loops, we can neither sense, hence observe, thereby collect a variety of information for the above process, nor decide as well as implement actions in accord with these processes. Or, put another way, without OODA Loops embracing all the above and without the Ability to get inside other OODA Loops (or other environments), we will find it impossible to comprehend, shape, adapt to, and in turn be shaped by an unfolding, evolving

reality that is uncertain, ever changing, and unpredictable.” (Boyd 2018, 383) 76 | EXPLORING DECISION ADVANTAGES In other words, it is the constant change of realities that necessitates adaptations. Further, the OODA loop is Boyd’s interpretation of what to do, and how to do it, to deal with the realities and uncertainties in the erratic nature of warfare. In an effort to understand the past, present, and future of warfare he continued to revise his work throughout his life. His theoretical contributions to this field were primarily realized through extensive briefings derived from his presentations and succinct essays (G. Hammond 2001; Boyd 2018) As a result, the legacy of his ideas has proliferated in a manner similar to Sun Tzu’s work; although relevant, it has fragmented into various components rather than forming a cohesive theoretical framework. As Osinga posits: “Reading through Boyd’s work nowadays one does not encounter novelty or experience difficulty following

his arguments and accepting his ideas. His language and logic, his ideas, terms and concepts are part and parcel now of the military conceptual frame of reference.” (Osinga 2007) Boyd’s continued relevance is evident in postmodern war studies focused on asymmetric warfare and in fourth-generation warfare (Osinga 2007). Additionally, recent NATO declarations 21 on multi-domain operation (MDO) emphasize the use of strategies for ‘outthinking’ and ‘outpacing’ adversaries. These are remnants of, and arguably not conceptually linked to the more comprehensive perspective that Boyd captured in his strategic theory. These fragments are still of value, offering important insights and opportunities for the creation of new theoretical assemblages and ways of thinking. Other researchers have tried to develop add-ons such as the dynamic loop (DOODA) by Brehmer (2005), or Breton and Rosseau (2005) having a cognitive C2-focus (C-OODA). However, the fact that contemporary military

organizations have absorbed Boyd’s ideas (and other thinkers who have turned them into new concepts) should not be perceived as any other than Boyd himself did when he analyzed his problem at hand and then synthesized the pieces into something new. They are examples of destruction and creation, much like his own approach (Boyd 1976). Despite these deviations and the emergence of new theoretical concepts, the OODAmodel 22 remains largely elusive to most readers. Instead, the model’s intrinsic ideas appears to be distorted and used in a fragmentary form. This may have to do with the complexity of the concepts integrated within the theory, despite the model’s apparent simplicity, especially the popularized version (Fig. 2) which illustrates a comparatively spartan and sequential process. Yet, Boyd never illustrated the loop in this way in any of his works (Richards 2020). Instead, and near the end of his life, 21 NATO, “Multi-Domain Operations”,

https://www.actnatoint/article/multi-domain-operations-enabling-nato-to-out-pace-and- out-think-its-adversaries/, accsessed 29th of January 2024. 22 In the context of Boyd’s work, the terms model and theory are used interchangeably. EXPLORING DECISION ADVANT AGES | 77 he produced the richer, and more complex version depicted in Figure 1. Similarly, he never claimed that going through the loop more quickly would provide the strategist or user with a decisive advantage (Richards 2020, 144). Figure 2 The OODA loop depicted as a simple sequential process (Richards 2020, 144). Boyd’s ambition to condense his core ideas into a simple schematic may lead readers to interpret his OODA loop solely as a model meant for fast decision-making cycles. As Osinga (2007) states, the “simplified version [of the OODA] tends towards an exclusive focus on speed of decision-making, while obscuring various other themes, theories and arguments that lie behind and are incorporated in it.” (2007,

7) This does not exclude any application for rapid decision cycles – they have their role to play as well. However, a more comprehensive investigation is necessary to understand the wider meaning and conception laid out within the complete loop (Fig.1), including the argument put forward by Richards that the model include more than one loop (2020, 155). The OODA loop is an open-ended nonlinear process utilizing feedback and feedforward channels to enable adaptation to both internal and external changes. Furthermore, it is constructed as an analytical tool that emphasizes context and a temporal dimension. Boyd promotes multifaceted interactions to gain the insights to make propositions to be acted upon and encourage short-cuts by implicit guidance and control (IG&C) where experience can co-create opportunities that can be seized. Having focused on the OODA loop as a whole, we will now move to the four components. 78 | EXPLORING DECISION ADVANT AGES The constituent parts of

the loop (Component level) Examining the constituent parts of the OODA at a component level (Fig.1) provides a deeper understanding of Boyd’s conception. According to Hammond, editor of ‘A Discourse on Winning and Losing’ (Boyd 2018), Boyd first used the word ‘sensing’ instead of ‘observation’ but then changed it for linguistic reasons (2018, 384). Sensing would arguably be the more appropriate contemporary term given that most of the information today is data retrieved by sensors as an input to the next stage. Nonetheless, what follows is a brief description from left to right of the four main steps and their respective links, chiefly based on Boyd’s compiled presentations (Boyd 2018) . The loop begins with Observation where the human user assesses the environment to include how it interacts with oneself or one’s own relation to it, the threats, and opportunities. It is the task of collecting all that matters or may matter The sensing or collection of relevant

information is fundamental as it updates the individual or the group on changes, new events, outcomes and enemy responses. Observation is therefore the primary source of new information. The feedback loops from a Decision (also referred to as Hypothesis (Fig.1)) and Action (also referred to as Test (Fig.1)) are interactions between the sources (inputs )and the empirics (outputs) It should be noted that Orientation shapes Observation through implicit guidance and control (IG&C). This can be interpreted as the known unknown requires detection This also means that the first stage interrelates and interacts with all of the other three. Importantly, the sensing from Observation provides a foundation of inputs from which Orientation, or what Boyd called ‘the big O, can commence’(2018, 384). Orientation shapes observation, decision, and action. It ensures internal alignment from a continuous interaction with the external environment. Logically, it therefore exerts power over what,

where and why we observe. Conversely, it is also shaped by observation. However, most observations would be meaningless without orientation (Osinga 2007, 230). Boyd’s description of Orientation is relatively detailed, and as seen in Figure 1, it consists of several components, including user’s genetic heritage, cultural traditions, previous experience, education and training, latest information and the analysis and synthesis that follows. These elements constitute a complex set of filters and inherent biases that shape any succeeding decision or action. If not observant to our own preconceptions and other subparts involved in our understanding and knowledge-building, which can lead to confirmation biases, ignorance or blindness that corrupts the process. He motivates these subparts by emphasizing that without these “we do not possess an implicit repertoire” of skills and experience, and that “without analysis and synthesis[]we cannot evolve new repertoires” to deal with the

unfamiliar (Boyd 2018, 383). Orientation also influences the way ahead. It produces various responses including EXPLORING DECISION ADVANT AGES | 79 insights, visions, focus and direction which subsequently leads to either a Decision on a preferred Action, or directly to Action via the IG&C (Fig. 1) This ‘shortcut’ is accentuated by Richards who stated that “our actions will flow from it implicitly, that is, without explicit[]commands or instructions, most of the time” (2020, 11). The Decision calculus is a hypothesis for testing - to see how the action of it shapes or becomes shaped. It appears to derive from the notion Boyd makes in other passages in his presentations about uncertainties being ever present. Changes within the environment are constant, and uncertainties ever present through unpredictability camouflaged in randomness. Consequently, assessments are essential. Boyd considered the requirement to assess “the accuracy and depth of common understanding in

an organization to be one of the primary functions of leadership.” (Richards 2020, 18) Accurate assessments and shared awareness do not only enable initiative via the IG&C link. It also deliberates adaptation to changing situations by, with and through the integration of feedback loops and the IG&C on both sides of Orientation. Now, from accurately assessing any situation one of two thing will be occurring. Either actions will be taken implicitly via the IG&C link, or via explicit decision-making. The former does not require a superior decision to be made, although some sort of decision-making will prompt the action. Decision is therefore perceived by Boyd as “a choice of a course of action, a trade-off in a wider trade space about some future state of affairs and their consequences[]It becomes a hypothesis to be tested by the Action” (Boyd 2018, 385). Situations requiring more detailed coordination in space and synchronization in time will be managed through

explicit Decisions. These decisions and following actions are perceived as unique, hence requiring explicit orders or instructions. As such, they are prone to hypotheses testing 23. Moreover, they represent an independent circular loop (observation, analysis and synthesis, hypothesis, and test) (Richards 2020, 20– 21). Surprisingly, Boyd discusses the action component more than he does decision (Osinga 2007, 232). Action either implements a Decision, or it is an execution via the IGC-link and based on experience that aligns with certain schema that call for action. Richards (2020) emphasizes that, to enable actions via IG&C-link, organizations should reach “excellence in [their] repertoire” without becoming predictive in such actions (Richards 2020, 19). Boyd defined ‘repertoire’ as accustomed or known actions within an organization that can readily be actioned upon should the situation arise (2020, 19). Since decisions represents hypotheses, they also signify double

loop learning, which after assessment and training could become part of the repertoire. According to Boyd, Actions should be rapid, surprising, and varied, therefore Actions and the Decisions must be fed back into the system for validity on the correctness of existing orientation patterns (Osinga 2007, 232). The result of 23 As visualized in figure 1, Boyd explicitly gives Decision a secondary meaning by adding hypothesis, and in an analogous way adding test to action. 80 | EXPLORING DECISION ADVANT AGES Action completes the loop and reconnect to new iterations that are assessed through new observations informing Orientation and so forth. A doctrinal framework for joint targeting This section aims to describe the foundation for joint targeting, as defined by NATO and the US. It discusses three key aspects: the joint targeting cycle (JTC); the two methods used (deliberate and dynamic), of which the dynamic method is the most relevant to this thesis; and lastly the intelligence

support activities that augments the subtasks within joint targeting. The relevant basic definitions of joint operations and joint targeting was defined in Chapter 2. The section is therefore a continuation and a more detailed description of the concept of joint targeting. Through the influence of a ‘joint approach’, both US (US DoD Joint Publication 2018a) and NATO doctrines for joint operations (NATO 2019) seek to improve effectiveness in major operations. Joint (or major) operations are planned, conducted, and sustained at the operational level of warfare to accomplish strategic objectives within theatres or areas of operations, by linking strategic objectives to tactical level operations (NATO 2019). Joint operations are conducted within a designated joint operations area (JOA), where armed forces are “deployed and employed in accordance with a strategy to achieve military-strategic objectives” to attain a desired end state (NATO 2019, 21). Consequently, the commander of a

joint (or multinational) force decides on how tactical activities are generated to achieve those strategic objectives. The key principles for ‘allied operations’ described in NATO AJP-3 are: unity of effort, concentration of force, economy of effort, freedom of action, definition of objectives, flexibility, initiative, offensive spirit, surprise, security, simplicity, and maintenance of morale (NATO 2019, 29–32). In addition, there are several operational considerations that may complicate the application of those principles and the use of military force. Although the doctrine describes the principles as a “coherent approach to complex and dynamic problems”, and that they “attract broad agreement as to their importance and relevance” (2019, 29), the design, organization, integration and conduct of joint operations appears to be highly complex. However, many other doctrines also support or are informed by the concept of joint operations, and scholarly critique has been

raised. In a critical interrogation, Finlan et al. argues that joint operations as a concept has evolved from an idea back in the 1980s, “into an assumption today or a “taken for granted” notion that exists within international and national military spaces without much critical reflection” (Finlan, Danielsson, and Lundqvist 2021, 7). They argue that the concept of joint operations has not fully matured knowledgeably, and as a result, military organizations and practitioners have encountered challenges in bridging gaps through translation, EXPLORING DECISION ADVANT AGES | 81 where “individual service culture” often eclipses the original notion and hinders the practical application of joint operations (2021, 14). Yet, the joint approach dominates the contemporary idea of warfare, whereby the elements of two or more services operate under a joint commander to accomplish common operational objectives (US DoD Joint Publication 2018a). This dominance is evident through the

incorporation of the joint targeting concept into western military organizations and their related doctrines (NATO 2016; NSO 2021; US DoD Joint Publication 2018b). Further, the concept of joint targeting intimately relates to comprehension, tempo, and high-quality decision-making. Joint targeting is an essential feature of the joint approach (Penney 2023). It concerns the process of selecting and prioritizing targets and matching the appropriate response to them, considering operational requirements and capabilities 24 . From a US perspective joint targeting resides within joint fires (Joint Chiefs of Staff 2019). Joint fires are the integrated and synchronized use of weapon systems or other actions “during the employment of forces from two or more components in coordinated action to produce desired effects in support of a common objective.” (2019, 19) The interrelation between joint targeting and joint fires is defined as “related capabilities and activities grouped together to

help commanders synchronize, integrate, and direct joint operations. The joint targeting process matches and integrates appropriate joint fires capabilities to validated targets to create desired effects and outcomes.” (2019, 10) Contemporary targeting has progressed along the axis of technological innovations and at the pace of air power’s evolution (see Chapter 2). It now entails a broader, more comprehensive approach, incorporating considerations and efforts conducted both before and during an operation, and at all levels of warfare. This approach entails efforts to integrate both lethal and non-lethal means against enemy combatants, military object, civilians participating in hostilities, or even stakeholders, neutral actors and civilian object through non-kinetic methods such as cyber operations. Central to its function is the ability to convey an intent with clear targeting objectives, and to conduct the planning, execution, and assessment of all targeting engagements. As

suggested in a comprehensive study on contemporary challenges to targeting, the results from targeting relies on four factors: ‘time’, ‘intelligence’, ‘competency‘ and the ‘cognitive ability’ to adapt to changes (Pratzner 2015, 82). However, the cognitive capacity to adapt promptly to changing situations requires considerable flexibility in the rapid allocation of forces and assets, as well as the competence and intelligence required to effectively utilize sensors and effectors as relevant to the specific situation at hand. 24 US Joint Publication 3-60. 82 | EXPLORING DECISION ADVANT AGES The six phases of the joint targeting cycle The joint targeting process is executed through an iterative cycle of six phases (Fig. 3). It is established and directed by the joint forces commander (JFC) Figure 3: The NATO version of the joint targeting cycle (JTC) as described in the latest Allied Joint Publication 3.9(B) (NSO 2021, Edition B:38) At the most abstract level, the

joint targeting cycle (or loop), is composed of six phases with similar content and aim in both US’s and NATO’s targeting doctrines 25. The process commences with the (1) Commander’s intent, objectives, and guidance, (2) Target development, (3) Capabilities analysis, (4) Commander’s decision, force planning and assignment, (5) Mission planning and force execution and (6) Assessment. The cycle is neither time-constrained nor time-dependent and steps may occur concurrently (U.S Joint Targeting School 2017, 95) The joint targeting process is characterized as a systematic approach employed by military forces to 25 Phillip R. Pratzner, “The Current Targeting Process,” in Targeting: The Challenges of Modern Warfare, ed PAL Ducheine, ML Schmitt, and F.PB Osinga ( TMC Asser Press 2015), 80 Pratzner illustrates how the two processes (or cycles) are substantially similar. EXPLORING DECISION ADVANT AGES | 83 integrate and synchronize targeting efforts across all operational

domains to achieve desired effects and objectives. A simplified interpretation of this cycle identifies four primary components: guidance, planning, execution, and assessment. However, such oversimplification may obscure the complexities and interconnections between intelligence and operations. Even if the doctrines state that the process is not time-constrained/ time-dependent, other relational factors are. As argued by Duchaine et al (2015) it is a comprehensive approach where combinations of multiple means are to be deconflicted, coordinated and synchronized. Further, Pratzner’s (2015) notes that the process neither is done in a void, nor that time is irrelevant. These activities are generally conducted in a dynamic, high-paced setting, where the adversary is advancing and maneuvering simultaneously in several domains and subject to critical time constraints. Others have the same arguments referencing the US JP 360 (Joint Targeting) doctrine as having major gaps between the

doctrine and any operational application, and failing to “connect its effects-based approach to the true rhythm of operations.”(Pinnix 2021, 72) They all agree that time, be it related to speed, pace or tempo, has to be accounted for in any targeting application. Taking a closer look at the separate phases of the joint targeting process, the first phase establishes the commander's objectives, targeting guidance, and intent for an operation, subject to any strategic constraints and guidance. It provides the overarching direction and priorities for the targeting effort and initiates a collaborative and multi-disciplinary staff work. The most vital collaboration is between intelligence and operations. The doctrines accentuate this collaborative approach as illustrated in Figure 4. 84 | EXPLORING DECISION ADVANT AGES Figure 4: The relationships between the JTC's phases and the intelligence products. Source: Joint Publication 3-60, Joint Targeting (Washington, DC: Joint

Staff, September 2018), II-6 The relationships between the JTC's phases and the intelligence products are clearly visualized in Figure 4. The proper selection of targets and the subsequent target intelligence material production stems from having a foundation built as illustrated. From a threat perspective, the intelligence function constructs this foundation by consolidating all relevant information and intelligence that affects the joint force within the operational environment but it is mainly concerned with the adversary. The joint intelligence preparation of the operational environment (JIPOE) is a product as well as an ongoing process of appreciating the situation and the environment. It informs a joint force in its planning of an operation, to include being an important input to phase 1 in the joint targeting cycle. The second phase, (target development) focuses on identifying, developing, and prioritizing targets. The intelligence function uses various target intelligence

products and targeting databases to fulfil the required level of precision and details. This phase is equally dependent on data, information, and intelligence of the adversary, to include intentions, capabilities, and any vulnerabilities within an adversary’s system. Furthermore, as described in the US instruction that prescribes the target development standards (US DoD 2016a), target development is “best performed with a clear understanding of the guidance” given by the Commander in phase one (2016a, 25). Typically, specific targeting objectives are derived from the Commander’s intent, objectives, and guidance. Once established, the targeting EXPLORING DECISION ADVANT AGES | 85 objectives drive the target system analysis (TSA) and target development (TD). The TSA is a product describing the ‘target system’ of the adversary in detail, from the top level down to individual targets. Each target consists of ‘target element’, which are the specific “features or

objects that enable the target to function” (2016a, 28). The detailed description of each target is further developed in electronic target folders (ETF) that are, in turn, managed via targeting databases. Each target considered for target engagement will be developed in such detail that their respective critical element(s) can be identified, including the aimpoint, or desired point of impact (DPI). The US instruction defines critical elements as an “element of the target that enables it to perform its primary function”, and an aimpoint (DPI) as a “point on the target designated for weapon impact or penetration.”(2016b, 28) The US instruction, CJCSI 3370.01B, provides some examples of what could constitute a critical element - for instance, a “critical element for an airfield could be the air traffic control tower” (US DoD 2016b, 28). These terms and their applicability are associated with standardized geodetic expressions derived via precise point mensuration (PPM) 26

and used in conjunction with targeting precision guided munitions (PGMs) and coordinate-seeking weapons (CSWs). The target system analysis (TSA) is an essential continuation of the JIPOE. It contributes to joint targeting by identifying relevant targets linked to the commander's objectives and priorities. These targets are nominated as potential targets for inclusion in the joint target list (JTL). The TSA allows the targeting staff to determine the military importance of each target and the potential effects that could be achieved by targeting specific components in the target system. Furthermore, the vulnerability analysis within a TSA, informs the third phase (capabilities analysis) to match appropriate capabilities and weapons to exploit those vulnerabilities and create desired effects. It also supports the effective prioritization of targets based on their importance, linkage to objectives, and potential for achieving desired effects when engaged, which allows for the

development of the prioritized joint integrated target list (JIPTL) in the fourth phase (Commander’s decision, force planning and assignment). Lastly, the TSA and ETFs all become vital instruments for assessing the effects post-engagement, due to their detailed information of each targets function and role within its larger context. In essence, TSA is a critical analytical process within the joint targeting cycle that systematically examines adversary systems, identifies high-payoff targets, determines their importance and vulnerabilities, and facilitates precise and effective targeting to achieve commander's objectives. This underscores the role of 26 The authoritive instruction for target coordinate mensuration (TCM) is CJCSI 3505.01E and applies to all services, headquarters and coalition partners “that conduct TCM using National Geospatial-intelligence Agency (NGA)-validated mensuration tools, methods, and geospatial-intelligence (GEO INT) sources during joint or

coalition operations”. See CJCSI 350501E , US Chairman Joint Chief of Staff: US DoD, 2022, 1. Mensuration refers to the process of measurement of a feature or location to determine an absolute latitude, longitude, and elevation to support the employment of coordinate-seeking weapons (CSWs). 86 | EXPLORING DECISION ADVANTAGES intelligence as essential to any successful application of targeting, as Pratzner (2015) asserts. However, as some research suggests, in time and resource constrained environments, more flexible approaches are required (Penney 2023). Others claim that the “targeting model needs to evolve, and as such the integration of intelligence that feeds that model must likewise evolve.” (Pinnix 2021, 73) This indicates that there is room for improvement, not only in the process as such, but also in how intelligence can deliver an efficient support to targeting. The third phase, (capability analysis), investigates and matches the most appropriate capability to use

for each target to determine the best way to create desired effects while minimizing collateral damage and complying with laws of armed conflict. This analysis concurrently becomes a precursor for which tactical forces and assets that are likely to be assigned to execute the targeting once it has been decided. The suggested capability matching is often referred to as ‘weapon-to-target assignment’ (Kline, Ahner, and Hill 2019), and has to be thoroughly examined to ensure that all resources are available, and then integrated, coordinated and synchronized with the tactical level considering their planning and the situation at hand. To support this, and other parts of the joint targeting cycle, designated targeting liaison personnel from the services should be augmenting the joint fires element (JFE) of the joint force, or any equivalent body authorized to oversee the joint targeting process as outlined in the US JFE manual (US Chairman Joint Chief of Staff 2021). Phase four

(Commander’s decision, force planning and assignment) is a recurring decision point of the targeting cycle, where the commander decides what to target, when to target and with what capability and for what purpose. The decision is made by the joint targeting coordination board (JTCB) based on the aforementioned analyzes and a list of options (NSO 2021) which in turn will affect force planning and assignments. As indicated by Duchaine etal (2015) targeting is a balance of resources, risks and efforts, where the commander can decide how to employ force more economically, using a cost-benefit approach to achieve the objective at the lowest affordable cost (2015, 3, 19, 203), or applying other key principles as more important. This is an internal opportunity to enforce or adjust the targeting decision policy. The board meeting (JTCB) aims to ensure swift and effective coordination of targeting efforts and to authorize a valid JIPTL. The vital considerations should be resolved before a

board meeting. The decision initiates the fifth phase (mission planning and force execution) which is conducted at the tactical level supported by assets retained within the joint commanders’ decision. At the tactical level, the targets are engaged using various methods and capabilities, still in accordance with JFC decision, targeting directives, and the defined legal framework of the operation. Since each component command (tactical level) has its own objectives derived from the joint operational plan, they are actually the entities that often nominates targets and suggests effects, assets, and weapons within working groups and at the JTCB. The JFE and the targeting process EXPLORING DECISION ADVANT AGES | 87 as such facilitates for the coordination and synchronization of all targeting engagements. Consequently, the tactical level has impact on the targeting conducted within a joint force. However, some targets based on their significance and the joint force commander’s

objectives and priorities, are likely to get a higher priority to which tactical resources will be directed. The cycle is completed through the conduct of the sixth phase (Assessments). Various assessments are made to accurately assess the effects originally sought within the JFC’s objectives, guidance, and intent. The last phase revisits all other phases of the cycle to accurately assess this, to include evaluating the effects created on each target (battle damage assessment (BDA)), the munitions effectiveness assessment (MEA) and the potential collateral damage assessment (CDA). The BDA is evaluated in three steps: first a physical assessment that determines whether the target was hit or not. Second, a functional assessment to assess how the function of the target was affected, which is subsequently followed by a more consolidated third step assessing the more long-term effects on the target system as a whole. In combination, these assessments inform future targeting efforts

through a cyclical process of re-evaluating objectives, developing new targets, and re-prioritizing based on the evolving situation. Software is an important aspect that supports the targeting process The joint targeting process is supported by software systems (Joint Targeting System (JTS) in NATO and Joint Targeting Tool (JTT) in US that are standardized tools for reasons of ‘interoperability’ (Zinca 2018, 25). The software systems support the storage and multi-level management of relevant data and information (to include ETFs, target lists, pictures, maps, link charts, assessment reports and targeting objectives). Moreover, they are partially automated, and designed to integrate adjacent command and control, decision-making and mission planning (Zinca 2018, 22–23). Targeting automation is the use of these computer systems, applications, and database technologies to speed the up targeting-related information processes (US DoD Joint Publication 2018b). However, contemporary

targeting automation does not integrate AI systems. The only indicated exception is the Israeli AI systems Gospel and Lavender that purportedly focus on individuals connected to the Hamas, or other armed terrorist networks (Davies, Mckernan, and Sabbagh 2023; McKernan and Davies 2024). Without any insights, or official presentations by the Israeli Defense Forces (IDF), media speculations are met with IDF denials. However, following the logics of this thesis exploration into AI applications for targeting purposes, the two systems (Gospel and Lavender) are likely based upon artificial intelligence (AI). As a hypothesis, they both fuse and cross-reference multiple layers of information from 88 | EXPLORING DECISION ADVANT AGES different datasets to generate suggestions for Targeteers (intelligence analysts working with targeting). Additionally, given the confined battlespace, the IDF uses the systems to recommend entities (probably focusing on individuals) that may qualify as a

military target, given a nomination process similar to the US and NATO equivalent. The datasets that Gospel and Lavender uses are presumed to be from intelligence collection, including satellite imagery, aircraft and drone imagery and video streams, cyber intelligence, cellphone intercepts, human intelligence, and open-source information. The multitude of data necessitates an integration of primarily computerized imagery analyses. The use of AI and neural networks to perform classifications can greatly reduce the burden and time to process the data for intelligence analysts. New software systems could be required to meet current challenges. As argued by Pinnix (2021), “for targeting to have maximum impact, there must be time to connect the dots of the broader network and leverage information generated through processes, which is a key weakness of dynamic targeting.” (2021, 73). Pinnix points to some important aspects relevant for this project His statement suggests that effective

targeting depends on understanding the broader operational or strategic context. Without sufficient time for analysis, dynamic targeting may reduce its effectiveness. Moreover, dynamic targeting often involves time-sensitive decisions. The term ‘key weakness’ indicates that the rapid pace of dynamic targeting may inherently conflict with the slower, deliberate analysis needed to achieve maximum impact. The potential trade-off between speed and effectiveness implies a need for improved software tools including AI applications and improved workflows within the methods used. The two methods of deliberate and dynamic targeting The doctrines provide guidance on using two methods when performing targeting at the joint level: deliberate and dynamic. Deliberate targeting applies when there is sufficient time to add the target to mission planning and schedule target engagement (NSO 2021). Deliberate targeting is more fixed than dynamic targeting It is often managed in episodes of 24h

cycles and, predominately for the air tasking purposes a planning period of 3-4 days (72-96h) from initial planning to execution. Dynamic targeting on the other hand is a more flexible method to be applied when the time for planning is limited. Specifically, this method can be used for targets that are either identified too late or not selected in time to be included in deliberate targeting, but when detected may be engaged. Dynamic targeting can also be applied for emergent targets of opportunity that appear in the battlespace. The more flexible dynamic targeting process is visualized in Fig. 5, and is termed F2T2E2A, an abbreviation of the seven stages within (FIND, FIX, TRACK, TARGET, ENGAGE, EXPLOIT, and ASSESS). EXPLORING DECISION ADVANT AGES | 89 Figure 5. The flexible F2T2E2A dynamic targeting process (NSO 2021, Edition B:89) Additionally, dynamic targeting methodology can be applied at the component level. NATO defines dynamic targeting as: Dynamic targeting engages targets

that due to the dynamic changes in operations, present a threat to force or mission, and whose criteria supports the commander's objectives[] Engaging these targets may be possible by redirecting existing assets in accordance with the Commander JTFs intent and targeting guidance. (NSO 2021, Edition B:36) This means that the dynamic targeting process handles emerging threats and opportunities 27 , all of which require an immediate responsiveness. These are all potential targets observed and detected by any collection asset within the joint intelligence network, and that may, or may not be on the joint prioritized target list (NSO 2021, Edition B:31). Following a procedure of identifying these emergent targets and confirming them as legitimate to prosecute, they are prioritized, which may challenge others targets already in the joint priority list. This is a complex process that consists of subtasks, which has been further analyzed and interpreted 27 The term emergent target is

also used in this dynamic context. 90 | EXPLORING DECISION ADVANTAGES from the doctrines used in this chapter. Figure 6 elucidates the vital subtasks, as well as the intelligence support activities. This interpretation is used to further derive the specific subtasks and intelligence support activities when defining the unit of observation in Chapter 4. Notions on the dynamic targeting method This section describes the project’s understanding of the dynamic targeting method (moving from left to right in Figure 6). An emergent target (potential target) is detected by an unmanned aerial vehicle (UAV) transmitting via real time link to headquarters (HQ). The emergent target now needs to be identified and classified to determine what it is and how it should be handled. The OUTPUT of phase 1 (FIND) is then to decide which classification the emergent target gets (a-e). This classification determines the INPUT of phase 2 (FIX) where class a and b are the most important. As seen in the

picture, a only requires confirming fact, whilst b needs to be validated as a target and then decided if it is a Time-Sensitive Target (TST), a High Pay-off Target (HPT) or High Value Target (HVT) and if there are any restrictions before it can be managed as a. The OUTPUT of phase 2 is a positive identification (PID), accurate location and a definition of its prioritization. Phase 3 (TRACK) a target may require sensors that are ‘keeping an eye’ of the target. In phase 4 (TARGET) there are two approaches. As an INPUT the target is either on the JIPTL, in which case the order may already be issued for engagement, or the target is not on the JIPTL. If it is not, the example considerations (a-g) in the green box in phase 4 needs to be determined and decided before JFC can issue orders, which are the OUTPUT. The last three phases (ENGAGE, EXPLOIT and ASSESS) are executed after the decision to execute a target engagement has been made. The most critical task within these three phases is

arguably the battle damage assessment within ASSESS. If the desired effect is met, then assets can be re-allocated to support other target engagements. If not, then a re-attack may be necessary, if the opportunity still exists EXPLORING DECISION ADVANT AGES | 91 Figure 6. The interpretation of the Dynamic targeting method used in this project to identify subtasks to the process. 92 | EXPLORING DECISION ADVANTAGES Intelligence support to targeting Intelligence support is paramount for effective targeting, providing key inputs to all phases within the joint targeting cycle (JTC). Further, it augments the Commander’s intent, objectives and targeting guidance during phase 1 by assessing the operational environment, analyzing, and identifying the vulnerabilities of the adversary, and subsequently develops measures of performance (MoP) and measures of effectiveness (MoE) to enable post-targeting assessment. Arguably, the most critical contribution of intelligence is to perform

target development, a process by which targets are identified, described, and characterized through conducting a target system analysis (TSA). The results are used to augment the prioritization between targets, which is continuously re-assessed during operation and combat assessments. NATO targeting doctrine (AJP 39) defines how various intelligence support activities augment the different phases of JTC (Fig. 7) Figure 7. The relations between the different phases in JTC and the key intelligence support activities being done (NSO 2021, Edition B:60) EXPLORING DECISION ADVANT AGES | 93 The most critical contribution of intelligence support activities is the analysis of the targets that creates the most desired effects on the adversary. It involves finding those targets and having the resources and opportunities to engage them. However, in reality this assessment is complicated. The problem is two-fold First, it relates to estimate and foreknowledge of the adversary and the

apparent difficulties to identify key vulnerabilities and performing a valid target system analysis that was brought to attentions by Pinnix (REF) . Second, it relates to time, efficient resource management, and being able to command and control the intelligence support of fast-paced, dynamic targeting. It may not always be that the slowest targeting method should define the pace. Finally, intelligence support activities augment both of the targeting methods. This necessitates priorities of intelligence collection assets. The intelligence function within a joint force will accommodate the appropriate management of intelligence collection operations. However, the dynamic targeting method manages emergent targets that could be unexpected which requires the collection operations management (COM) to be fully attuned and capable of shifting focus - re-directing collection assets to address these targets and opportunities. The concept of the intelligent agent The concept of artificial

intelligent agents is central to this thesis. According to renowned AI scholars Russell and Norvig, an intelligent agent is "something that perceives and acts" (Russell and Norvig 2014, 34). As Russell (2019, 43) explains, the intelligent agent is defined by three factors: the nature of the environment in which the agent operates; the agent's connections to the environment; and the agent's objectives. In Chapters 5 and 6, the respective AI models are intelligent agents that act within their respective environments, processing perceptual inputs and converting them into actions based on the specific conditions and objectives. An agent’s functionality must be evaluated based on its capacity to meet these goals within the constraints of its environment and programming. Depending on the complexity of the problem and the environment, to include “fully observable vs. partially observable”, “single agent vs. multi-agent, and “deterministic vs stochastic” the

sophistication of the agent’s design will vary (Russell and Norvig 2014, 42–43). The concept of intelligence itself is important to understanding intelligent agents. This project adopts Russell's (2019) characterization of human and machine intelligence as "Humans are intelligent to the extent that our actions can be expected to achieve our objectives," and machine intelligence as "Machines are intelligent to the extent that their actions can be expected to achieve their objectives" (Russell 2019, 9). Other voices such as philosophers Arkoudas and Bringsjord (2014, 34) suggest that AI comprises “artifacts” capable of displaying behaviors that humans consider intelligent. Essentially, these definitions emphasize action as the ultimate purpose 94 | EXPLORING DECISION ADVANTAGES of perceiving, reasoning, and learning. This is measured by the achievement of defined objectives. The justification for integrating intelligent agents into warfighting

concepts depends on their ability to perform tasks or generate actionable insights that enhance efficiency or decision-making. Two assessments further motivate the importance of intelligent agents in this thesis. First, the integration of AI in modern warfare is becoming an intellectual necessity driven by advancements in related research fields (Horowitz et al. 2018) and strategic military initiatives such as the command and control-initiative “JADC2” (Department of Defense 2022) as well as real-world alleged instances such as the Israel Defense Forces’ reported use of AI for targeting in the Israel-Gaza conflict (Davies, Mckernan, and Sabbagh 2023). Second, as weapons and sensor systems at the tactical edge grow more sophisticated, higher echelons will increasingly rely on automation, autonomy, and collaboration with intelligent agents to maintain efficient decision-making, command and control (C2) arrangements, and joint operational concepts (Hoadley and Lucas 2018). Meeting

the challenges of dynamic and fragmented battlespaces therefore requires new paradigms and innovative solutions. AI provides tools to gain decision advantages by enabling faster and more informed decisions than adversaries. This necessitates intelligent agents designed to process relevant data, achieve specific objectives, and augment human decision-making processes, particularly in joint targeting. What distinguishes AI from other technologies is its capacity to emulate human-like intelligence. There exists methods and applications for “simulating, extending, and expanding human intelligence.” (Jiang et al 2022, 1) Whether or not AI systems achieve self-awareness, they are at least anticipated to rival, if not surpass, human intelligence in specific domains (Russell 2019). However, the development of AI is evolving at a high speed and is becoming increasingly more complex in structure. Jiang et al (2022) suggests that this development makes it harder for humans to comprehend and

follow decisions and argues that “In the long term, AI systems and human beings need to establish common cognitions in terms of the outlook on the world, life, and values.” (2022, 12) They indicate the need to build a bridge “to connect the cyber and the physical worlds” to “synchronize, correct, and update” the former from the latter (Ibid. 2022, 12) Trust therefore becomes an important issue between humans and machines. Currently, An organization’s reliance on its own models is contingent upon trusting the data they are based on. Therefore, it is essential to establish the criteria that determine when recommendations from AI models can or cannot serve as a valid basis for decisionmaking, as well as “under what circumstances recommendations from AI-models can or cannot justify actions.” (Bovet Emanuel 2024, 188–89) This concludes the description of the three concepts that constitute the theoretical framework. The next sub-section will provide certain critical

reflections after which the chapter offers its conclusion. EXPLORING DECISION ADVANT AGES | 95 Reflections Criticism and defense of Boyd Boyd’s controversial thinking and his legacy still appears as a fertile ground for criticism Robinson (2021) argues in his book The Blind Strategist that Boyd’s ideas on maneuver warfare is “deeply flawed” as he was relying on fraudulent accounts of the German Blitzkrieg by Wehrmacht veterans, thereby “fatally tainted [the maneuver warfare concept] by this information.”(2021, 19) However, western militaries still seem to utilize the concept of maneuver warfare as a way to approach land warfare. The ways each state has incorporated it is, however, contextual, and therefore mutated from the original ideas crafted by Boyd and others. Robinson’s criticism also refers to Boyd’s “preconceived notion of providing the universal applicability of the O-O-D-A loop” and that Boyd is “forever trapped in the feedback loop of his closed

system.” (2021, 20) Countering such criticism can only be done by an analysis of the comprehensive structure of the OODA loop. Boyd deliberately built-in resilience against tendencies to seek out and prefer information that supports pre-existing beliefs (often referred to as confirmation bias). The loop, if studied and used as laid out in Boyd’s theory, is arguably the recipe for how to avoid getting trapped in, or becoming isolated, in one’s own closed system. Boyd praises interaction with the environment – not isolation. Other sources of criticism have been related to two viewpoints. First, how to classify Boyd’s work (as a strategy, or more of a general theory of conflict) or as a conceptual foundation for maneuver warfare (Osinga 2007, 5). Second, how to interpret or make use of the simple and easy-to-comprehend OODA loop (Ibid. 2007, 6) Though the criticism originates from two separate assessments on Boyd’s work, the former being a classification issue and the later an

implementation issue, they arguably share a common ground. They miss the point The discourse between the reader’s own mental schemata, preconceptions and not least problem at hand, and the interpretations made when reading Boyd’s ideas. According to Osinga, any interpretation of Boyd’s work “must be informed by his methodology[]and go through the same learning process [as Boyd]” (2007, 8). The origins and strands that informed Boyd, and gave birth to his insights must therefore come in a rather condensed form starting with Boyd’s own arguments about strategic thinking disclosed in one of his early presentations titled A Discourse on Winning and Losing: 96 | EXPLORING DECISION ADVANT AGES the theme that weaves its way through this ‘Discourse on Winning and Losing’ is not so much contained within each of the five sections, per se, that make up the ‘Discourse;’ rather, it is the kind of thinking that both lies behind and makes-up its very essence. For the

interested, a careful examination will reveal that the increasingly abstract discussion surfaces a process of reaching across many perspectives; pulling each and everyone apart (analysis), all the while intuitively looking for those parts of the disassembled perspectives which naturally interconnect with one another to form a higher order, more general elaboration (synthesis) of what is taking place. As a result, the process not only creates the ‘Discourse’ but it also represents the key to evolve the tactics, strategies, goals, unifying themes, etc. that permit us to actively shape and adapt to the unfolding we are a part of, live in, and feed on. (Osinga 2007, 8) Undoubtedly, the written legacy of Boyd’s presentations challenges most readers. His ambition to refine, upgrade and perfect his work created many versions which may be a source of confusion as to his intentions. Regarded as one of the most important military strategic theorists of the twentieth century (Gray 1999,

90–91), Boyd’s propositions within his OODA loop will serve as the theoretical underpinning of the empirical outcome from the experimentation. The four components of the OODA loop also frame the concept of targeting, both the more wide-ranging joint targeting process, and the tailored method labelled dynamic targeting. Furthermore, Boyd’s military theory is rooted deeply in the discipline of war studies. The combination of a long-established theory tailored towards decision-making / C2 in a military domain, and the military doctrines addressing the central aspects of targeting offers a foundation for concept exploration and theory development. It also creates a bridge to the technologies of AI/ML and the subsequent models that are to augment decision-making within the joint targeting process. Boyd was heavily influenced by his extensive study of patterns through the history of warfare. The OODA loop provides a foundation for decision-making It supports the project’s aim and

objectives. The attention Boyd pays to Orientation and the significance he stresses on adaptation along with his two optional routes to action makes the theory both vibrant and challenging for any military organization trying to apply it to the full extent. His perception of time became synonymous with his concept of tempo in warfare that conflates time with speed. Tempo is conditioned by a common intent within a joint force, which he combines with a reformistic bottom-up C2 perspective. EXPLORING DECISION ADVANT AGES | 97 According to Boyd, the essence of warfare is human perception, not the weapons. He suggests that “Machines don’t fight wars. Terrain doesn’t fight wars Humans fight wars. You must get into the minds of humans That’s where the battles are won” (G Hammond 2001, 122–23). Whilst this may still hold true, the question is for how long? The relationship with machines is no longer the same as it was 50 years ago. Now the machines work without the human. Using

drones as an example, these machines fly by themselves carrying munition or sensors with them and dispatching relevant information on the progression of the battle to humans and returns to base by themselves if programmed to do so. In reference to the citation of Boyd, at least the first sentence is open to debate, evident from contemporary drone wars in Nagorno-Karabakh (Eckel 2020), the alleged use of AI by Israel in Gaza (Davies, Mckernan, and Sabbagh 2023), and the acclaimed AI-weapons used in the war in Ukraine (Russell 2023). Although this can be perceived as an argument in opposition to the theory, it is not. Boyd’s OODA loop is still valid The fact that AI has entered just makes the approach to agency in warfare more heterogeneous. It refines some of the stages now that artificial intelligence is becoming more sophisticated, and now that we (humans) are more reliant than ever on data as the input of observation, computers to orient ourselves and provision for decision

support, and machines that acts and perceptually experiences distant actions. It appears clear why Boyd defined Orientation as the big ‘O,’ and why he paid so much attention to it. It is the ‘brain’ in the model, shaping the perceptions of individuals and groups within, and guides or direct all the other three stages implicitly or explicitly. the estimation and foreknowledge are congregated by Boyd into Orientation. The outcome of each iteration of Orientation is arguably the equivalent of the commonly used term common operational picture (COP) and achieving a situational awareness. COP is defined as “a single identical display of relevant information shared by more than one command that facilitates collaborative planning and assists all echelons to achieve situational awareness” (Chairman of the Joint Chiefs of Staff 2021, 42). It is also a contest for learning, and of constantly improving one’s situational awareness. If properly done, orientation will be the foundation

for adaptation. The accentuation towards implicit over explicit is to “gain a favorable mismatch in friction and timefor superiority in shaping and adapting to circumstances” (Richards 2020, 12). On the one hand, this can be viewed as a leverage for efficient command, enabling rapid actions directly from Orientation via the implicit guidance and control (IG&C) link, as Richards describes (2020). Certainly, actions taken directly from superior situational awareness and experience (or ‘fast transients,’ as Boyd calls them), can be beneficial within the C2 of functions to include targeting. However, in the light of contemporary challenges in targeting (Ducheine, Schmitt, and Osinga 2015), the current practices of C2 within targeting would have to be conceptually reengineered. Specifically, the IG&C link between Orientation (sensing, estimate & foreknowledge) and Action indicates that Boyd understands and promotes networked, self-regulated organizational structures that

98 | EXPLORING DECISION ADVANTAGES more recent researchers suggests (Kott and Alberts 2017). This will likely necessitate more advanced technology to reengineer the sensor-to-shooter concepts. Moreover, ensuring situational awareness requires the organization as a whole to assert a shared outlook, including a superior intent that unifies and promotes subordinate initiatives via such an IG&C link. This is arguably challenging and requires relevant means of communication, training, doctrines, and cohesion. Additionally, joint targeting situations may arise that necessitate compliance with certain policies, rules, and regulations, potentially annulling the employment of IG&C. Reflections on the theoretical framework and other approaches Building a theoretical framework that support the project should facilitate novel perceptions and ways of actors and actions within the joint targeting practices. Based on the research problem and its question, the focus is on AI as an

intelligent agent and actor. As a consequence, this narrowed down the number of potential theories. In the end, the theory of Boyd’s and his OODA loop was considered suitable to allow for new ways of dealing with the challenges and activities it contains. Other frameworks could potentially have been employed for this project. Pending their respective framework of thoughts and concepts, this would have affected the scope of this thesis. At one point, a theory that acknowledged the recursive relationship between human-machine was considered, since the AI-applications are thought to function in a collaborative environment. Their function can, from a practice theory perspective, be perceived as an integral part of targeting practice. As defined by Schatzki (2012), the commonalities of practice theory is the idea of socially organized constellation where the activities are understood from how they are regulated by rules, procedures or other types of guidance. Two problems emerge from

Schatzki’s notions. The first relates to how practices evolve as it must allow for changes to the practices since AI will intervene to improve the function. The second problem is that the theory must allow for non-human (AI) agency. Schatzki’s (2013) theoretical concept does include support for evolving practices – it embraces the emergence, persistence and dissolution of practices. Accordingly, technological innovations are, using Schatzi’s conception, entities that affects practices and may cause changes to how a practice is performed (Loscher, Splitter, and Seidl 2019). Yet, the second problem seems outside of the conception of practice theory, despite, for instance, Pickering’s (2005) claims for a mutual constitution of material and human agency. Though changes within practice theories are likely to follow from this other theories have conceptions that incorporate the problems addressed here. The actor-network theory (ANT) offers a relevant perspective on how we observe

and study new practices. ANT is distinctive because, as Law (1992)informs us, “it insists that networks are materially heterogeneous”, and contends that all entities EXPLORING DECISION ADVANT AGES | 99 should be analyzed in the same terms (1992, 379). Hence, in this view, “the task of sociology is to characterize the ways in which material join together to generate themselves and reproduce institutional and organizational patterns in the networks of the social” (J. Law 1992, 379) In other words, the theory departs from the more conventional currents of social science and suggests that everything is of equal value and connected to each other in some way. Most importantly, according to Callon is that “ANT is based on no stable theory of the actor; rather it assumes the radical indeterminacy of the actor”, which has opened the social sciences to non-humans (2007, 273). This indeterminacy of the actor has created criticism towards the theory, as being too tolerant and

relativistic. Other leading figures such as Mol that ANT is a conceptual set of methodological tools to use when investigating a phenomena within a practice in order to “make specific, surprising, so far unspoken events and situations visible, audible, sensible. It seeks to shift our understanding and to attune to reality differently.” (Mol 2010, 255) Moreover, there is no quest for cause or a causal explanation, the aim is rather to trace effects (Mol 2010). In sum, the methodology behind ANT would have been supportive of this thesis if the scope would have been solely towards the non-human agency and if the aim were to interpret how their participation created a more heterogenous workforce in targeting practices. It would have required a constructivist approach ANT does not provide the theoretical framework that is needed to situate the experiment on AI and targeting. It does not support answering the research question For that, a theory of decision-making and doctrines on

targeting are needed. Yet, ANT (and related theories) could still be useful and valid in conceptualizing and thinking about AI and the human-AI relationship. The recursive relations between humans and machines Investigating recursive relations between humans and non-humans (in this context AI being perceived as intelligent agents) are relevant as the progression of AI moves towards these becoming context-aware multi-agent systems (Du et al. 2024) However, the influence of intelligent machines has been discussed for decades (DeLanda 1991). Agents that are extant at the same level of existence will form a heterogeneous agency. But to support an investigation into this we need tools that can be operationalized in order to enable a thorough understanding of such a phenomena. Of relevance is what, and how the actor acts and interrelates to other actors in a network. Mol states that “actors associate with other actors, thus forming a network in which they are all made into “actors” as

the associations allow each of them to act. Actors are enacted, enabled, and adapted by their associates while in their turn enacting, enabling and adapting these.” (Mol 2010, 260) Agency tends to become what socio-technical researchers attributes to the concept and what practitioners make of it. George and Bennett (2005) contends that human agents 100 | EXPLORING DECISION ADVANT AGES and social- and material structures are “mutually constitutive,” (2005, 129–31)and argues that an agent-centered intentional change is unique to human agents. This is accentuating a post-modernistic critique of the positivist tradition in the philosophy of science. On the other hand, theories like actor-network-theory (ANT) broaden the agency to include non-humans and maintain a position where intentionality and reflexivity is less important. This enables a wider social science perspective Moreover, as argued by Law (1992) we need to distinguish between ethics and sociology when studying

practices where non-human entities are existing on the same ontological level as humans. Yet, he claims, they are not identical: “To say that there is no fundamental difference between people and objects is an analytical stance, not an ethical position. And to say this does not mean that we have to treat the people in our lives as machines. We don’t have to deny them the rights, duties, or responsibilities that we usually accord to people. Indeed, we might use it to sharpen ethical questions about the special character of the human effect” (J. Law 1992, 383) Additionally, inquiries associated in human-machine collaborations are arguably flawed because only one side gets a voice. Human interpretation is the solitary perspective. Investigations in these matters may require a cultural-hermeneutic approach (Ben-Ari and Levy 2014) and interpretivist methods. The experimenters’ own observations as well as interviews with human participants are the two main data collection methods in

which such a cultural-hermeneutic approach can be operationalized. Nevertheless, research on recursive relations between intelligent agents, both humans and non-humans, is becoming more relevant also within war studies and in military organizations as humans moves towards “human-machine teaming” (HMT) (Nurkin and Siegel 2023, 5). The integration as such is perhaps a lesser concern than how AI is employed and for what purposes. Conclusion The chapter has explained and discussed the theoretical framework involving three concepts: Boyd’s strategic theory, with a particular focus on his OODA loop; and the joint targeting doctrines and associated instructions released by NATO and the US; and the concept of intelligent agents. Boyd’s theory is not only descriptive, but it also aims to inform the practice and policy of how military decision-making ought to be structured and performed. It has a prescriptive purpose to serve as a normative guide for rapid and adaptive decision-making.

Understanding this duality of his theory enhances its application. Through a descriptive lens it provides a framework for understanding how humans and its organizations process information and make EXPLORING DECISION ADVANT AGES | 101 decisions in dynamic environments. It captures the complexities of human cognition and behavior and highlights the necessity of learning and adapting. It underscores the competitive nature of decision-making. Through a prescriptive lens the theory suggests how to approach complexities and uncertainties. It points to agility, speed, and adaptation claiming that an advantage can be gained by operating at a faster tempo than the adversary. The doctrines on joint targeting are perhaps less controversial but at least as comprehensive as the theory. However, the chapter have sought to provide a thorough account of what joint targeting is while selectively narrowing the focus to the most critical aspects that directly support the subsequent experiments. The

most vital aspects are the method of dynamic targeting along with the intelligence support activities. These will be further analyzed when the theoretical framework is operationalized in the Methods chapter. Importantly, the project employs Boyd’s theory to situate the experiments within the broader context of military decision-making, emphasizing adaptability and learning. Doctrines provide a baseline condition for the experiments by defining contemporary joint targeting practices, while Boyd’s prescriptive lens highlights how military organizations can better structure OODA loops to enhance agility and adaptability. Together, the theory and doctrines frame the problem and establish a foundation for applying AI as a treatment to improve joint targeting. The AI applications are referred to as intelligent agents, utilizing the concept laid forward by Russell and Norvig (2014). The integration of three concepts supports the experimental setup and its simulation experiments, offering

a pathway to enhance human decision-making and organizational adaptability in dynamic environments. Furthermore, all three concepts stress the importance of contextual understanding. They depend on an understanding of the broader context to function and to be effective. This is applied in Boyd’s OODA loop accounting for both internal and external changing conditions. As for the joint targeting process, not least the dynamic method, this is evident in identifying how specific targets relate to the larger network of adversary capabilities and objectives. For intelligent agents (AI), it is inherent in the definition of the concept. Moreover, the current development in AI focuses on expanding its context awareness. Dynamic targeting involves rapid decision-making processes to achieve maximum impact. The trade-off between speed and effectiveness seems to be an existing challenge. It highlights a tension in military operations between agility and the depth of intelligence analysis. The

need for speed could result in prioritizing short-term gains over long-term impact. Notwithstanding this issue that may persist, AIapplications could augment the dynamic method and its supporting intelligence activities by generating the information required to make informed targeting decisions in time. AI applications could accelerate the analysis of complex 102 | EXPLORING DECISION ADVANT AGES adversarial networks and provide actionable insights in near real-time. This may require improved workflows and an altered C2 structure that can synthesize and leverage information rapidly across a targeting enterprise. The chapter also reflected on Boyd’s theory, including critiques and alternative theoretical approaches. Academic criticism is not inherently negative; rather, it highlights interest in the subject. Boyd’s military theory, deeply rooted in war studies, combined with military doctrines on targeting, provides a robust foundation for concept exploration and potential

theory development. This combination also bridges AI/ML technologies, and the two models are designed to augment decisionmaking in the joint targeting process. While alternative theoretical approaches, such as ANT, do not offer the necessary framework to situate the AI and targeting experiment or address the research question, they remain valuable for conceptualizing the human-AI recursive relationship. With these conclusions, the thesis now turns to the project’s design and methodology. EXPLORING DECISION ADVANT AGES | 103 104 | EXPLORING DECISION ADVANTAGES Chapter 4 – Method Introduction This chapter sets out the theoretical outlook that informs the choice of research methods within the project. The research is situated at the heart of war studies through its investigation of how joint targeting can be affected by an integration of artificial intelligence to augment related decision-making. The choice of observation is motivated by the desire to extend existing

research frontiers by generating new knowledge and intellectual concepts. It engages with contemporary problems in joint targeting and synthesizing novel solutions using artificial intelligence within a dynamic targeting process. The research mirrors Boyd’s approach of nurturing new conceptions and innovation – “to improve our capacity for independent action” (Osinga 2007, 132). It analyzes and synthesizes scholarly work and official documents “across a variety of domains or across a variety of competing independent channels of information[to]evolve a new repertoire to deal with unfamiliar phenomena” (Boyd 2018, 383). Its exploratory approach also provides the context for the two major sections in the chapter: the design of the study, and the application of research methods. The project uses a variety of methods ranging from qualitative analyses of the OODA loop and targeting doctrines to survey and experimental testing using different modelling and simulations. This

multimethod approach facilitates addressing the research question and provides a systematic account for how AI can augment decisionmaking within the joint targeting process. It therefore imposes an interdisciplinary approach of mixing military practical knowledge and theory with heterogeneous 28 (or diverse) technological knowledge and the empirical AI-produced knowledge from the experiments. The approach is vital as it is the combining, or integration, of diverse pieces of knowledge that creates the understanding necessary to solve the problem at hand. The approach corresponds to the authoritative study on military power by Stephen Biddle (2010) to explain how material and nonmaterial factors interact to produce combat outcomes. It also utilizes a mixed method approach, including historiography, case method, formal theory (or the use of mathematical language to describe relations) and simulation experiments to facilitate inference from observations to “conjectures about the

future”(2010, 9–10) . As with this project, 28 In this context heterogeneous technological knowledge is referred to as diverse and varied types of information that could come from different sources, domains, or perspectives. For instance, knowledge gained from an academic course at a university can be perceived as tacit knowledge, online knowledge accessed via github.com as explicit knowledge, and knowledge gained by discussing with other researchers or during experimentation as embedded knowledge. It is a form of knowledge integration, see for instance Drucker ‘The New Society of Organizations, Harvard Business Review, no. September-October 1992, 95–104 EXPLORING DECISION ADVANT AGES | 105 Biddle used simulation experiments to address “soft variables like force employment in a rigorous way” and thereby “compensating for the weaknesses of individual methods taken alone.”(2010, preface) This also anchors the project within War Studies. As an outcome of the method

section follows a brief introduction of the two experiments and an operationalization of an analytical tool that ties the theoretical framework with the results from the two experiments. The chapter closes with reflection and conclusion. Philosophical and theoretical perspectives 29 related to the research topic This research is located philosophically and theoretically within the social sciences. The research has a cross-disciplinary approach consisting of perceptions of technological dimensions, philosophical human-machine perspectives, warfare practices, mathematics, and computer science in order to analyze the problem space and to synthesize the relevant parts that constitutes a potential solution space worth exploring. Second, it is built upon a theoretical foundation of both a decision-making theory and relevant military doctrines on joint targeting practices. Third, it seeks to look beyond current paradigms and beyond present capabilities towards novel and ostensibly weak

domains of knowledge. It searches for insights by exploiting opportunities from external technological trends for military purpose. Furthermore, it perceives the two AI-models as artefacts capable of displaying behaviors that are considered to be intelligent (Arkoudas and Bringsjord 2014, 34). Employing the concept of ‘intelligent agents’ (Russell 2019, 42) defines the research as one that relates to these agents as actors, thereby investigating features such as their intelligent behavior displayed under controlled and well-defined conditions. The question of what indicates the presence of intelligence has historically been the subject of debate, challenged by what Arkoudas and Bringsjord (2014, 46) describe as “three principal philosophical criticisms”. These include: (1) Dreyfus's critique from the 1970s, largely dismissed as invalid but notable for highlighting the importance of capacities like imagination; (2) Block's critique of machine functionalism, introduced

through his China brain thought experiment in the late 1970s; and (3) Searle’s Chinese room thought experiment from the 1980s (Ibid., 2014, 47, 49). Notably, these criticisms focus on strong AI, in contrast to the narrow (or weak) AI applied in this project (Ibid. 2014, 47,49) Notably, these criticisms focus on strong AI, in contrast to the narrow (or weak) AI applied in this project. 29 A theoretical perspective describes the philosophical stance of informing and determining the research methodology. For more on this topic, see Michael Crotty, The foundations of social research: Meaning and perspective in the research process (London, Sage Publications Inc., 1998) 106 | EXPLORING DECISION ADVANT AGES However, the very idea of intelligent machines is closely related to rationality and the attempt to ‘mechanizing reasoning’ (Arkoudas and Bringsjord 2014, 36). The ability to make logical inferences, to extract conclusions from premises lies at the heart of AI. It is something

that AI does efficiently and can be automated Notably, there are levels of complexity in artificial reasoning. The most difficult reasoning problem is that of ‘conjecture generation’ (Ibid. 2014, 38), which relates to AI having an ability to generate interesting solutions that logically follow the premises provided. Even if this could be debated from a philosophical standpoint, evidence from recent medical research show that this ability is already in place. With machine learning and generative AI, medical laboratories have demonstrated that complex models can be built automatically 30, and that these AI-models can generate novel solutions to existing problems 31. Since this thesis is concerned with applications of AI, rather than the study of human thought, it is less concerned with the internal workings of human cognition than it is of constructing problems that can be solved by intelligent agents in support of decision-making within joint targeting practices. The project also

represents a pragmatic and innovative approach, that dovetails with the assertion of former US Secretary of Defence, Donald Rumsfeld, about force transformation, to “change not only the capabilities at our disposal, but also how we think about war” (2002, 29). The research question narrows the scope towards how intelligent agents can be applied to augment a military process involving the abilities to understand, decide and take action in a timely manner. Given the pragmatic approach, it naturally resonates with the philosophy of pragmatism. It declares that “the reality exists in the world, and it supports the objective nature of science” (AlAbabneh 2020, 81), allowing multiple explanations and interpretations for social science. Consequently, this means that acceptable knowledge can be provided from either, or both, “observable phenomena and subjective meaningdependent on the research question” (Al-Ababneh 2020, 82–83). Moreover, Al-Ababneh argues that pragmatism

“focuses on practical applied research, integrating different perspectives to help interpret the data” (2020, 83). Hence, this philosophy is situated between the positivist and the interpretivist research philosophy. It also means that the research can have a cross-disciplinary approach. 30 Heaven, Will Douglas, MIT Technology report, 15 February 2023, “AI is dreaming up drugs that no one has ever seen”, https://www.technologyreviewcom/2023/02/15/1067904/ai-automation-drug-development/, accessed 27 August 2024 31 Trafton, Ann, MIT News Office, 25 May, 2023, “Using AI, scientists find a drug that could combat drug-resistant infections”, https://news.mitedu/2023/using-ai-scientists-combat-drug-resistant-infections-0525, accessed 27 August 2024 EXPLORING DECISION ADVANT AGES | 107 Design Design is the strategy or methodology of how the research is performed, thereby directing what methods that will be used. According to Al-Ababneh, methodology “provides a rationale

for the choice of methods and the particular forms in which the methods are employed “ (2020, 77). Building on this, Ritchie et al (2014) highlight that design addresses several key practical aspects, including defining research questions, selecting data collection methods, and managing time, finances, and other methodological considerations. They emphasize that design is closely tied to ‘good planning,’ a continuous process of review and adjustment that spans most stages of a project (Ritchie et al., 2014, p 48) On the other hand, Swedberg (2012, 2014), a sociologist and proponent for theorizing, argues against much of the mainstream sociology that is focusing more on methods rather than originality (2012, 7–8). Instead he promotes creativity in social science, using the “context of discovery” as “an independent element in the research process”(2012, 7). Swedberg’s (2012) way of theorizing falls largely in line with the approach in this project. Finally, the

experimental nature of this thesis necessitates a mixed-methods approach, aligning with IR-theorist Lawson (2015) concept of an “eclectic methodological frameworkwhich is better suited to the task of studying complex social and political phenomena” (2015, 18). The methodology used in this project The project’s methodology consists of two parts: an early prestudy and a main study. The prestudy seeks to generate ideas from real-world problems and to explore avenues of approaches to conceptualize specific contemporary targeting problems and to solve these with AI in a simulation environment. It intends to support the development of an innovative theoretical approach to the topic and explore the available data sources as well as the ways in which data could be synthetically produced. It fits well within the context of discovery (Swedberg 2012, 3) The main study begins by framing the research problem, defining the research question, and deciding what theories or theoretical framework

that would be most feasible to support the scope, aim and objective of the thesis. The results of the prestudy make it clear that the most interesting real-world problem relates to decision-making within a dynamic targeting setting. The purpose of the main study is to operationalize the research question by marrying it in a logical way with the theoretical framework to support the empirical work of the thesis. It can be viewed as the context of justification, a process which Swedberg attributes as “the form in which thinking processes are communicated to other persons” (2012, 3) 32. 32 Swedberg is referencing the work of Reichenbach in the citation used here. 108 | EXPLORING DECISION ADVANT AGES The principal stages of the prestudy The prestudy includes five principal stages, each briefly explained here: - Observation and literature review This stage involved discussions with subject matter experts in AI at the Swedish Defence Research Agency, at Linköping University, at

SAAB, and one of the leading commercial research centers (IBM AI research center in Zurich) during the first two years. These consultations supported the framing of the phenomenon (AI) and gave valuable inputs to the interrelations between a human problem and how this can be turned into a problem that could be solvable by AI. It supported the internal process of thinking and theorizing. Parallel to the more in-person observations, a rather extensive literature review was conducted. The main themes of it, and the end result can be found in Chapter 2. - Formulating the problem, research gap, and the central concepts This stage was more of a continuous process throughout the prestudy, than a followon stage to the previous. The outcome of this is in Chapter 1 - Real-world inferences The linkages with the Swedish Armed Forces (SwAF) have been important. Early on in the research process I established contacts with SMEs within the SwAF. It was driven by ensuring that the internal

theorizing, including the early conceptualizations of potential use-cases and problems resonated well from a realworld practitioners’ point of view. The main challenge at this stage was not to get carried away by additional ideas stemming from these discussions. The real-world inferences that were established during this period enabled a continuous interaction throughout the research process. It created opportunities and occasions to communicate the research, not only to the SwAF, but to NATO, and to industry partners and universities. These opportunities were important for the research external relevance and validity. The specific inferences are discussed in Chapters 5 and 6. - Exploring the sources of data and avenues of approaches. As Swedberg elegantly frames it by citing a famous conversation in A Scandal in Bohemia’ book by Sir Arthur Conan Doyle “I have no data yet. It would be a capital mistake to theorize before one has data. Insensibly one begins to twist facts to

suit theories, instead of theories to suit facts.” (2012, 12) Retrieving data to investigate potential venues and feasible use-cases proved more challenging than anticipated. This was partly due to information security concerns and partly because of a longstanding tradition of not preserving or readily sharing data from exercises and operations. Consequently, the experiment conducted in Chapter 6, relating to the use of optimization algorithms, led to the development of the project’s own datasets. To EXPLORING DECISION ADVANT AGES | 109 ensure the correctness (external validity) of the data used in this experiment, subject matter experts from the SwAF were used to verify it at an early stage. As for the experiment involving the use of neural network (Chapter 5), this idea was conceptualized after being able to retrieve open access data from a NASA repository containing high-resolution satellite images. - Setting up a modelling and simulation environment Setting up the

simulation environment involves investing in computers that could perform the intended simulations. It also encompasses initiating collaborations with two of the Swedish Defence University’s strategic industry partners: IBM and SAAB, to enable a channel of technical interaction concerning each of the two subsequent experiments. The support from the two companies is defined in the Methods section and each environment implicitly displayed in each empirical chapter. The principal stages of the main study The main study includes six principal stages, each briefly explained here, and more thoroughly discussed in the next section (Method) and in each of the two empirical chapters (Chapters 5 and 6). Robinson (2017), a renowned expert in the field of modelling and simulation is used as the primary source for the methodological approach to the experiments conducted. - Operationalization Operationalization involves defining the research question, deciding on the theoretical framework to

use, and how to turn the research question into an empirical project to find the answer. The most critical part turned out to be how to employ a theoretical filter that would work on both experiments and for both Boyd’s OODA loop and the joint targeting process as well as the subordinate tasks (dynamic targeting method) and activities (intelligence support). The end result of this operationalization turned out to be not just a helpful tool, but a resourceful instrument that enabled the transition between the results and the theoretical framework. - Modelling Modelling involves the thinking and framing of the model, often referred to as conceptualization. Each of the two models have their individual presentations and considerations laid out in their respective chapters. What is more important to point out here are the lengthy processes of deciding on what specific problem to use as a case that is also feasible for what type of AI technique and method. The result of this stage will

shape the three subsequent stages that, in turn, becomes consequential for the overall results and conclusions of the project. - Building 110 | EXPLORING DECISION ADVANTAGES At this point in the process, the focus shifted to constructing the two models and refining them through iterative loops with simulations. The iterations were multiple for each of the models. Time was a key factor in the main study Each of the two experiments were preplanned to take six months from building to evaluation. Minor adjustments and complementary validations extended the intended timeline by almost 50%. - Simulation Through repeated simulations, errors were identified and corrected, leading to a robust and reliable representation of the problem through well-defined conditions in a controlled environment. - Evaluation Evaluation is vital to modelling and simulation. In this project it involved internal verifications and validations to ensure that the model actually did what it was supposed to

do, measuring its performance against pre-defined criteria, and exploring the possibilities of validating its external relevance with latent end-user. - Writing up the results and conclusions This stage includes the factual results and implications of the experiments. It also refers back to the theoretical framework and method used. Additionally, the implications of empirical findings have some extensions towards future research avenues. Experiments in social sciences Experiments in social sciences are nothing new, but they have many benefits. A persuasive argument for using experiments in social sciences is offered by Druckman (2022). He suggests that: Experiments are a central methodology in the social sciences. Scholars from every discipline regularly turn to experiments. Practitioners rely on experimental evidence in evaluating social programs, policies, institutions, and information provision. The last decade has seen a fundamental shift in experimental social science thanks

not only to their emergence as a primary methodology in many disciplines, but also to technological advances and evolving sociological norms (e.g, open science) (Druckman 2022, 6) EXPLORING DECISION ADVANT AGES | 111 Other researchers also advocate experiments with different, yet similar justifications. 33 Boulanin and Verbruggen (2017), for example, promote building on discoveries from previous research to improve existing, or develop new, knowledge. This is echoing other theorists within the field (Radder 2009; Swedberg 2012), not least Biddle’s (2004) authoritive work which combines simulation experiments with other contrasting methods into a ‘methodological triangulation’ (2004, 9). Shadish et al. (2002) argue that experiment offers a controlled environment, “in which an intervention is deliberately introduced to observe its effect.” (2002, 12) This can be understood as a way of exploring novel and complementary solutions to a defined problem. An experimental method

involving humans could also refer to a method of dividing individuals into different groups using randomization, leaving one group as a control group and then exposing the remaining groups to some sort of stimuli, measuring the outcome to better understand causal relations (Teigen 2014, 229). However, the experiments conducted in this project do not involve humans and aim to explore new ways of solving contemporary problems. This makes it difficult to compare the results to a control group, though the causal effects are easier to define, as they are directly attributed to the interventionnamely, the AI models. For armed forces around the world, experiments are considered especially important. According to the US DoD guidebook on research and engineering asserts that “experimentation fuels the discovery and creation of knowledge and leads to the development and improvement of products, processes, systems, and organizations” (US Department of Defense 2019, 2). Moreover, experiments

can be perceived as an overarching strategy for modelling and simulation (M&S) used within operations research (OR). The OR approach dates back to before World War II when military services turned to academia to help develop scientific approaches which aimed at solving military problems, to include the need to allocate scarce resources within operations in an effective manner (Hillier and Lieberman 2021). The OR literature defines its methodological characteristic as an attempt to search for a best solution (or optimal) to a defined problem (Quttineh and Larsson 2014; Y. Wang, Xin, and Chen 2022). Considering these perceptions from experts within social sciences (Druckman 2022), war studies (S. Biddle 2010), and the OR-field (Hillier and Lieberman 2021), they all engage in an experimental approach to find solutions to defined problems in controlled environments. Computer-based experiments are a logical next step in social science research. Furthermore, the reasons behind having

computer-based experiments as the main method for data collection are found in two methodological considerations prone to the project. First, the project has an exploratory and innovative approach. Social sciences literature on the subject defines this way as a dynamic view where the discoveries matters the most (L. Cohen, Lawrence, and 33 There are also several military practical handbooks on the matter. See for instance UK MoD, ‘Defence Experimentation for Force Development Handbook’, UK MoD, 2021. 112 | EXPLORING DECISION ADVANTAGES Morrison 2011), and it relates to an environments that can ‘facilitate exploration and discoveries’ to produce what Radder call ‘experimental knowledge’ (2009, 1). Since AI is intended to solve problems that will augment human decision-making, this constitutes the outline of using computers and intelligent agents to emulate the intended augmentation. Moreover, exploratory approaches are not controversial in social science and war

studies. Theorizing implicates creativity, a context where “the work of an experimenter,is to make vague intuitions explicit and testable.” (Webster and Sell, 2014, 43). Second, and based on the intention to use AI-models to solve human problems, the project requires controlled environments, specific conditions, and empirically established inferences. Simulations or computer-based experiments are often considered analogous to laboratory experiments, as both allow researchers to isolate and focus on specific phenomena. These methods have been extensively employed in military studies (Hill and Miller 2017), to explain military power (S. Biddle 2010) and, perhaps more importantly, to explore how AI can enhance military capabilities and optimize force employments (Zhao et al. 2019; Li et al. 2023; Hendrickson et al 2023) Likewise, it provisions for an ability to ensure that, what Boyd refers to as “internal consistency to match up with reality” (Osinga 2007, 134). By

contextualizing the experiments using the factual doctrinal guidance on targeting, it strengthens the reliability of the project since doctrines describe its intended employment. As a final note on justifications, and in relations to the paucity of methodological guidance covering AI research, it is worth highlighting that the specific subfield of AI in social science is still emerging. One of the few published methodological works, Machine Learning for Experiments in the Social Sciences by Green and White (2023), explores the use and application of machine learning and AI for experiments in the social sciences. However, their two methods were not applicable for this project Nevertheless, it demonstrates that AI for research is an evolving field in the social sciences. Guidance towards the conduct of experimentation using modelling and simulation The conduct of experimentation using modelling and simulation in this project necessitates the synthesis of relevant information from other

fields. Two of the most relevant sources for this are found in the field of Operations Research (OR), with Hillier and Lieberman (2021) and Robinson (2017) authoritive works. On the military side, the conduct of experimentation for research are often guided by best practices (US Department of Defense 2019; UK MoD 2021). These handbooks largely correlate with the guidance given in more scholarly work linked to social sciences (Webster and Sell 2014). The correlation may well be due to the latter replicating the former where it fits. However, apart from a basic association to the principles of EXPLORING DECISION ADVANT AGES | 113 experimentation: or the ability to create a controlled, artificial environment that allows for isolation and observations of the intervention or phenomena, there are more reasons motivating this project to be led by OR guidance. These are more explicitly referring to the conduct of modelling and simulation. OR is applied to problems that concerns how to

conduct and coordinate the operations or activities within an organization (Hillier and Lieberman 2021, 3). OR strives for efficiency and to aid in decision-making on all levels of an organization. It uses a scientific method characterized by first understanding the problem by gathering all relevant data, then constructing a model in an attempt to abstract the essence to represent the essential features of the situation or conditions, so that the conclusions (or results) obtained from the model are valid also for the real problem (Ibid. 2021, 3) Recent developments in artificial intelligence and machine learning, are increasingly becoming important tools of operations research (Hillier and Lieberman 2021, 11). The benefits of modelling and simulation – experimenting with models Modelling and simulations are essential tools when conducting the experiments in this project. There are benefits of experimenting with models as opposed to case studies because there are few examples of AI

in contemporary warfare. Moreover, and as suggested by Biddle (2010), “any ex post facto method – whether large or small n – faces a problem of selection on wars when testing theories of capability.”(2010, 10). The outcomes of interest to this project (performance of AI) can only be validated in the real world during wartime. However, the potential of AI-applications and its abilities to augment decision-making that concerns joint targeting-related problems can be done in simulation experiments. It demands constructing a model that captures the essential features of the problem , including the conditions that are necessary for solving it. The intelligent agent’s (AI-model) perception and actions can be controlled and observed, while holding all other aspects constant, in a similar way to how Biddle constructs his experiments (2010, 11). The input data will vary between the two experiments and so will the output. Both AI-models’ actions and output are simulations of how each

AI-model would act if it were integrated as a decision support system within a military targeting enterprise. The alternative pathway would have been to conduct it as a laboratory experiment with human staff and decision-makers, but it would have led to additional concerns, including infrastructural issues and ethical concerns with regard to humans as participants. However, it could be a logical next step (a future research pathway) from the simulation and evaluation of the models and preferred more than a fielded experiment that involves dynamic targeting. Williams (2013), an authority in operations research, motivates building models in this context as: 114 | EXPLORING DECISION ADVANTAGES “Experimentation is possible with a model, whereas it is often not possible or desirable to experiment with the object being modelled. It would clearly be politically difficult, as well as undesirable, to experiment with unconventional economic measures in a country if there were a high

probability of disastrous failure. The pursuit of such courageous experiments would be more (though not perhaps totally) acceptable on a mathematical model.” (2013, 4) Experiments with models using the scientific method described by Hillier and Lieberman (2021, 3), provide valuable insights without the risks and constraints associated with humans or with field experiments. Additionally, other benefits to experimenting with models can also be found. Building non-human intelligent agents that augment human decision-making within a targeting context, actually cocreates tangible applications that can be used to increase the understanding of social concerns, or the ethical and legal aspects of applying AI into targeting decision support. It warrants the study of other important, yet potentially sensitive topics, to include ongoing, deliberative attempts to codify ‘meaningful human control’ (Bode and Watts 2021, 4), or research that investigates how applied AI within decisionmaking

makes norms (Bode and Huelss 2024). Models, such as the two built in this project, could be used in different scenarios or for other use-cases, given their implicit limitations. Apart from their intentional employment as agents augmenting targeting-related decision-making, the models could be used to analyze exercises, hypothetical scenarios or even recreating historical events related to targeting. Specific scenarios that involve ethical dilemmas or strategies in targeting would also be possible, enabling both leadership training and the application of specific policies including frameworks such as the Laws of War and rules of engagement. The main contribution from the models used in this project is that they can facilitate a systematic study of the ways AI can be applied to complex military decision-making processes. Limitations to simulation experiments The most apparent limitation of simulations is the fact that the models are theoretical construction, or artefacts, and “not a

clear window that reveals nature directly to us” (Shadish, W. R, Cook, T D, Campbell 2002, 29) As a consequence, the results could be hypothetical, even fallible if based on too many unstated theoretical assumptions. Another limitation is that models are conditioned The are inherently context-dependent, which makes the results potentially irrelevant outside its defined conditions and context. Field experiments could be perceived as better environment to test and evaluate a model. They do, however, prerequisite a deeper understanding of the interdependencies in between different factors than that of an EXPLORING DECISION ADVANT AGES | 115 isolated laboratory experiment. Given Williams (2013, 4) motivations, and the potential for other interfering factors co-existing in such an environment, there are reasons to stay embedded in a more cost-effective, controlled environment, that can ensure observations and facilitate for unsolicited measures of performance of the artificial agents

working to augment any succeeding decision-making. As a supplementary discussion on limitations to simulation experiments, it should be noted that different AI techniques and methods bring different limitations to the table. If large language models (LLMs) are used, then the researcher must be aware of its level of context-awareness. The ability to have a nuanced understanding of context, such as its level of common-sense reasoning, and its ability to apply its knowledge flexibly to novel situations will be important concerns to consider. LLMs are built to generate an answer to a question or problem, not necessarily the correct or even the most rational one. Generating incorrect information (often referred to as hallucinations) is therefore an associated concern to deal with if LLMs are used. LLM-based models are also prone to perpetuate and amplify biases present in their training data leading to skewed, incomplete, or even inaccurate and discriminatory outputs. Pending the

complexity of the model, especially models that include neural networks, these models’ behavior, including internal inferences leading to their output or conclusion could make it difficult to explain and understand (often referred to as black boxes). Method The research method refers to the techniques and procedures including the various complementary activities to collect and analyze data based on the research question. The chapter confines the attention to the choice of use-cases, the concept of a model, and the way modelling and simulation is employed. It also addresses the technical expert support. Choosing models and use-cases One of the early decisions in the main study was to choose if one or two models were to be used, and the use-cases in which they should be applied. As previously defined, use-cases are scenarios, or well-defined situations and circumstances that are conditioned. The use-case approach is a practical and systematic method through a test-driven design

(Jacobson, Spence, and Kerr, 2016). The foreknowledge in targeting from the author’s side was extensive enough to support a swift period of defining and selecting use-cases relevant to the project. This foreknowledge also made the first step of modelling straightforward. However, the technical expertise in building the AI-models was inadequate, which meant that education and specific 116 | EXPLORING DECISION ADVANT AGES technological support had to be integrated in the project. After completing the prestudy, a decision was made to use two models, motivated mainly by the reasoning that two separate experiments will enable twice the amount of data collection, and a richer insight into the research question. The use-cases and models were selected upon three criteria: first, their ability to ensure compatibility and enable comparisons towards the OODA loop and the joint targeting cycle (JTC); second, the complexity of the task; and third, the required techniques. Five areas of

application were considered, all of which met the first criterium: 1. 2. 3. 4. 5. applications that process data and information from diverse sources and produce recommendations to facilitate efficient decision-making, applications that assist various stages of target development, applications that assist faster prosecution of emergent targets, applications that assist in optimization of the sensors and effectors, applications that assist combat assessments. The second criterium, the complexity of the task, represents a general difficulty when developing AI-applications. It is often implicitly interrelated to the complexity of the “task environment”, the level of precision in the “problem formulation” (or task specification), and the extent to which the environment is predictable and can be modelled in advance (Russell and Norvig 2014, 40,65). The more abstract the task environment is, the harder it is to accurately formulate the specific problem and the solution.

However, this can be mitigated through assumptions and simplifications, which is discussed in the next section. As a natural consequence of the second criterium, some use-cases were considered too difficult to model. It became quite apparent that this criterium was the major factor behind exclusions of applications and use-cases. Moreover, due to the complexities of the task, this had a secondary effect on the required data. Simply put, this criterium resulted in further adjustment and became the key factors in deciding what problems to model and simulate. The last criterium was simple. It had to do with ensuring that two different techniques were used to provide a broader foundation for the conclusion of the project, a deeper understanding of the ways in which AI can be applied to augment decision-making and providing a more thorough answer to the research question. In the end, one model engaged in a problem relating to area 3 and 4 using optimization algorithms (technique), whilst

the other model engaged a problem relating to area 2 and 4 using deep learning through neural networks (technique). EXPLORING DECISION ADVANT AGES | 117 Models as simplified representations of reality A model is an artefact and a simplified representation of the phenomena and the surrounding environment it imitates and subsequently simulates. A general aim is to build the simplest model possible, given the objective of the study, as suggested by Robinson (2017, 2745), and accrediting this aim as being faster, more flexible and easier to interpret. The methods of modelling and simulation vary depending on the problem and techniques used. Since this project includes two different problems and two different techniques, it still strives to simplify the overall approach. It draws upon the work of Hillier and Lieberman’s (2021) scientific method of OR, Robinson’s (2017) concept of modelling, Williams’s (2013) model building for mathematical programming, and Lundgren et al. (2010)

approach to optimization, to assist this process. The project roughly defines four basic stages applicable to both experiments. These are: (i) conceptual modelling based on a real problem description and intended solution, (ii) model building, (iii) simulation, and (iv) evaluation. This supports the outline of each experiment As the third and fourth phase (simulation and evaluation) are quite different and are therefore discussed in their separate chapters. Instead, we turn the attention to the first two stages Conceptual modelling and model building Robinson suggests that conceptual modelling is “the most difficult, least understood, but probably the most important activity” (2017, 2740). Consequently, the first phase is given more description and explanation before advancing to the subsequent phase. There is a broad agreement towards standard requirements on models amongst theorists (Stewart Robinson 2017; Hillier and Lieberman 2021; Williams 2013), that it should be

‘valid’, ‘credible’, ‘feasible’ and have ‘utility’ (Stewart Robinson 2017, 2744–45). Robinson explains that a valid model “should produce sufficiently accurate results for the purpose at hand”; that its credibility is related to the users’ perception; that feasibility is intimately connected to “the constraints of the available data and time”; and that utility is judged upon the models relevance (2017, 2744–45). Overarching these four requirement, is the requisite “to build the simplest model possible to meet the objectives” (2017, 2745). However, as suggested by Law, another expert in the field, a model can “only be an approximation to the actual system, no matter how much time and money is spent on model building Indeed, a model is supposed to be an abstraction and simplification of reality.” (A M Law 2002, 1283) Still, approximations can and should be accounted for. This is done by carefully accounting for the assumptions and simplifications

made. Indeed, the real problem has to be reduced to a simplified version in a way that still makes the results valid and reliable to a certain degree. The project uses the term ‘real problem’ to describe the environment in which the 118 | EXPLORING DECISION ADVANTAGES actual use-case resides, and the term ‘simplified problem’ to describe the environment in which the AI-model interacts. Assumptions and simplifications to the real problem The reduction of complexities and other less relevant factors are done through assumptions and simplifications. The project uses Robinson’s definition of assumption as a facet of limited knowledge or presumptions made when there are uncertainties about the real system, whilst simplification is a facet of the desire to create simple models and reduce unwarranted complexities (2017, 2741). Following this logic, conceptual modelling is the abstraction from the part of the real world it is representing (‘the real system’), and under which

the ‘real system’ may, or may not, currently exist (Stewart Robinson 2017, 2741). Understanding the real problem is essential, yet not the most vital. Robinson explains that “we rarely have the luxury of vast quantities of either knowledge or time”, and that even in an environment of immense knowledge and time, “a simpler model is often sufficient to address the problem at hand” (Ibid. 2017, 2740) Hence, the most vital part is to determine the level of abstraction at which the problem is addressed. Abstraction implies simplifications though still representative of the real problem. As all models are simplifications, they all include assumptions and disregards, or ignore some details. It becomes an iterative (repetitive) process to develop the necessary assumptions and simplifications. Although the conceptual phase is non-software specific, the decision to have neural networks in the first experiment (Model 1, Chapter 5), and optimization algorithms in second experiment

(Model 2, Chapter 6), made the process easier. The specific assumptions and simplifications for each experiment were revisited during the respective model building, primarily to reduce the complexities in each model. Throughout most of the second phase (model building) in both cases, the project maintained a balance between creating a realistic setting and avoiding excessive complexity. Meeting the requirements of models through external collaboration To ensure meeting the requirements put forward by Robinson (2017), especially in relation to validity, credibility and utility of the models, military subject matter experts from the SwAF were consulted to confirm the used input data, to define the real problem, and to justify the tactics used for ground-based surface to air missile defense (Model 1) and weaponeering solutions (Model 2). The interaction involved direct collaboration with the Air Force joint center of excellence (LSS), the Artillery Combat Training School (ArtSS), and the

Air Defense Combat Training School (LvSS). EXPLORING DECISION ADVANT AGES | 119 The collaboration with the Air Force (LSS) was mainly related to their understanding of targeting and the current challenges. Their statements were in line with the literature on the subject which added value by verifying the challenges. The discussions also supported the understanding of the threat from surface to air missiles, which confirmed the relevance of idea behind Model 1. Furthermore, the experts could also support the development of the data sets (weaponeering solutions) used for Model 2. It sharpened the data by providing valuable facts towards the feasibility and effectiveness of weapons and effectors. The interaction with ArtSS included a full day visit, and subsequent information exchanges with their experts. The objective of this collaboration was to investigate if there could be a feasible use-case for using deep learning and to understand the possibilities / limitations of current and

potential effectors and munition. The usecase which involved deep learning became a case in which the LvSS became involved. However, the support from ArtSS nevertheless was crucial to Model 2, since its effectors (see Chapter 6) included surface to surface munitions, in which long range artillery are perceived as joint weapons or weapons that can be utilized by other services for targeting purposes. LvSS, the Air Defence Combat School in Sweden, came to be the agency that was most frequently used in support of the project. This was mainly due to the development and subsequent validations of the results. The conceptualization as well as the model building required an understanding of the general tactics from both sides: the defense side, and the attacking or target side. The collaboration included multiple visits to LvSS and even involved a survey at the end of the experiment to support the evaluation and finetuning of the model. This is elaborated in more details in Chapter 5. The

collaboration with the SwAF allowed for an interaction where knowledge from professionals could be provided prior to the experiments to include confirmations to the quality of data input and a deeper understanding of how the real problem could be reduced to a simplified problem. The collaboration also provided valuable insights to enhance the accuracy and credibility in the actual simulations. Furthermore, it gave opportunities to present the intended solutions and exchange thought on how the AI systems potentially could be employed and from these exchanges assess their particular relevance and utility (Alberts and Hayes 2003; Stewart Robinson 2017). 120 | EXPLORING DECISION ADVANT AGES Strategic partnership and technical expertise Based on two basic strategic agreements between the Swedish Defence University (SEDU) and its commercial partners IBM and SAAB, the project could be partially sponsored. Both of the commercial partners involved agreed to pay for their own engagements in

the project. The establishment of this was finalized during the prestudy and became an enabler for the main study. As for their different roles and support of the project, both related to technical advice and support in the creation of the two models after the conceptualization was done. This made the two arrangements more straightforward as to what kind of technical advice and expertise were needed. The technical support from IBM was minor yet included two parts. After discussing the use-case with IBM research center in Zurich, I was appointed an advisor from IBM in Stuttgart. Secondly, it involved academic access to their software program where the optimization model could be built, including a cloud environment. After I had built the model, the technical expert appointed verified the CPLEX Optimizer programming (opl-coding) of the models, which were used for simulation and evaluation. The IBM advisor also proof-read Chapter 6 from a technical perspective, including validations of

mathematical formulations and the correlations with the solver’s (CPLEX) solutions. It was also a valuable step in the evaluation of Model 2 The model was later used as the main feature in a new collaboration with an IBM team during 2024, as a proof of concept (PoC), based on a request from the SwAF to provide a demonstration of a decision support tool for targeting purposes. The PoC was held in Stockholm on the 9 October 2024, and is perceived as another valuable indication of the models validity, credibility, feasibility, and utility (the standard requirements, see Stewart Robinson 2017, 2744–45). The technical support from the appointed expert at SAAB related to Model 1 specifically focused on the building of the deep learning algorithm. During the development of this model, no cloud environment was created. Instead, we worked collaboratively on two different computers, with meetings on the progression on a monthly basis during the year of completion. The model was built and

simulated as a joint effort, which is reflected in the patent application 34 that was handed in during 2023. The conceptualization, the evaluation, and the compilation of the chapter was done solely by the author. The expert from SAAB proof-read Chapter 5 from a technical perspective after its completion. The collaboration with IBM and SAAB ensured that the progression of the experimentation could be conducted with technical support. However, reflecting on the collaborative processes throughout this period, it is apparent that the responsibility for initiating, guiding, and sustaining the interactions was with the 34 The patent application no. 2300078-9, 20 September 2023 states two names as the inventors, the author of this thesis and Ella Olsson, SAAB Aeronautics. EXPLORING DECISION ADVANT AGES | 121 doctoral candidate. While these collaborations were essential, in view of reliance on the technical expertise and resources that the partners possessed, the process was often

challenging. The burden of initiative was with the project and required continuous outreach and strategic coordination to keep momentum. Nonetheless, fostering and preserving positive relationships with these actors was crucial. The nature of interdependencies required a level of adaptability to navigate varying levels of engagement and motivation across collaborators. This added complexity to the role of coordination as it was essential to maintain a constructive and collaborative atmosphere, even at times when there was a lack of reciprocity in the efforts with the main effect falling on the doctoral project. Maintaining this balance, ensuring the project advanced while managing mainly my own expectations and sustaining goodwill with the partners, proved to be an ongoing and demanding task. It highlighted the labor of relational management, where both technical reliance and interpersonal considerations intersect in collaborative academic and technical work. In the end, the

preparations, to include conceptualizations, education, training and attending courses in the subject, took more than two years, whilst the two experiments took around six months per experiment. A need for comparison? Experiments are comparative by nature (Webster and Sell 2014), because there is some sort of comparisons between a baseline condition and a treatment (the innovation). In relation to this project, a baseline condition is found in the current practices of joint targeting. As there may be several variations to current practices, the most appropriate description of the baseline condition are the two joint targeting doctrines discussed in Chapter 3. The treatment (or innovation) are the two AI-models, which are already defined as the unit of observation (UoO). However, what can be expected from the experimentation, where the observed intervention of AI-augmentation (treatment) is conducted for a specific portion of the targeting process and as a way of making the

decision-making more efficient, faster, or precise? Is a comparison necessary or possible? As the real problem is a problem or challenges from the real-world, the treatment is less prone to be compared to any baseline condition. However, this has been managed in the following way Each experiment is measured by its own performance. Each AI-model is given a set of criteria that will allow for measurements of its performance. As each problem is reduced to a simplified problem, the performance of each model, as in any other simulation, is conditioned. These conditions are the settings and the AI-models’ environment for the simulation. As both treatments are intended to solve an existing problem within joint targeting, any success in solving the problem is perceived as positive, meaning the treatment (or innovation) was effective in comparison to the current practices. However, this can and is in fact further explored through the 122 | EXPLORING DECISION ADVANT AGES performance

criteria for each experiment. These criteria attempt to describe in what ways the treatment is effective, to include measuring the AI-models’ precision, efficiency, time-consumption, and consistency. Validity Validity is an “important key to effective research” and thus a requirement for all research traditions (L. Cohen, Lawrence, and Morrison 2011, 179) According to Ritchie et al. (2014), validity is the extent “to which a finding is well-founded and accurately reflects the phenomenon being studied” (2014, 354). In other words, validity determines whether the project measures what it intended to, which in the project are the two experiments (Unit of Observation). The findings from the two experiments should reflect the phenomena, which is the AI-models performing the intended augmentation. If the results from each of the experiments reflect an augmentation and given that the experiments are set up correctly, then according to Ritchie et al, this would demonstrate validity.

Validity can be divided into different subparts, each reflecting a certain aspect of validity. Three are commonly applied to quantitative research: “measurement validity - relating to the degree to which the measures successfully captures the concepts they are intended to capture.”; secondly, “internal validity - the extent to which causal statements are supported”; and thirdly, “external validity - the extent to which the study’s findings can be generalised” (Ritchie et al. 2014, 356) However, concerns of validity may differ pending the field of research, and the way validity is operationalized is more vital than the term used to capture it. Validity was operationalized in three main ways: first, the precision in which it defines the real problem and the specific conditions it defines the reduction of the real problem into a simplified problem. The simplified problem, including its assumptions and simplifications defines conditions under which the experiment is

conducted. The results from each experiment are valid to the extent of the conditions under which it was conducted.; second, external AI-experts from IBM and SAAB have reviewed (proof-read) each chapter, in which they were technical support to verify and validate the findings. Furthermore, use of the SwAF experts is also considered valuable to establish validity; and third, results have been presented at several international conferences, the last of four presentations this year alone being the 33rd European Conference on Operational Research (EURO 2024). These conferences have provided opportunities to receive critical questions about the construction of the experiments, the model building and the algorithms used as well as other related issues for each individual experiment. All considered, this operationalization of validity is arguably supportive of all three components of validity (Ritchie et al. 2014) EXPLORING DECISION ADVANT AGES | 123 Reliability Reliability concerns

“the replicability of research findings” (Ritchie et al. 2014, 355) It means that if the findings can be verified when repeated, then the research has an elevated level of reliability. This can be easier in quantitative studies, given the access to the data and a well-defined approach, than in qualitative research which requires other forms to percept more elusive knowledge that could support reliability (Golafshani 2003; Ritchie et al. 2014) The experimentations could be referred to as being quantitative. However, it is a multi-method approach as the experimentations are partly built upon qualitative analyses. This includes the operationalization of the theoretical framework (Table 2), parts of the performance measures, and the interpretation and analyses of the interrelationship to the OODA loop as well as the JTC. Indeed, even the conclusions are interpretation of the findings. Notwithstanding this, the main issue of reliability is that it implies replicability and consistency.

This was operationalized in three ways: first, transparency – Model 1 has its training data defined as based on U.S Geological Survey and Space Shuttle Radar Topography Mission (SRTM) elevation data. The full system description of the model is defined in Chapter 5, along with the specific software used, the training environment, and the results. The same applies for Model 2, which has all the mathematical formulations define, the data set and examples of the output to be replicated and compared. The specifics of the software used, including the version is also provided. Indeed, transparency has been vital for the project to ensure that it can be replicable. Second, repeated tests – Model 1, being a neural network learns by iterations of training and fine-tuning. After the model was behaving the way it was intended to, it was trained extensively (see Chapter 5) on satellite images and then validated on new topographical data to show that it had learned its lessons. The model works

As for Model 2, being an optimizer that uses an exact method though a solver (CPLEX) developed by IBM and recommended by leading scholars (Hillier and Lieberman 2021), it requires no tests for consistency. After sensitivity analysis was confirming that the instructions (algorithms) were correct and that it could perceive the input and act by giving recommended attack options as an output, its abilities could be repeated without any deviation. Given the same input, and using the model as thoroughly defined mathematically in Chapter 6, it will give the same solution, every single time. All considered, this operationalization of reliability is arguably sufficient to support any external control of consistency of the two models. To summarize, the experiments were designed to ensure reliability and validity in both the empirical results and the overall project. The experiments are reproducible and replicable. In fact, the entire process of first conceptualizing the respective model, then

defining the AI-techniques to be used to solve the problem, and the data used, up to and including sensitivity analyses and verifications are all part of the 124 | EXPLORING DECISION ADVANTAGES design and described in each of the two chapters. The use of multiple re-runs with various samples and treatments receiving the same results also accommodates for the alleged reliability. Additionally, the validity is seen as an integral and vital part of the evaluation of each experimentation. By consistently referring to the ‘reality,’ or unit of analysis, within the experiments, the results can be logically inferred to both the OODA loop and the joint targeting process at both a system level and the relevant component. This iteration to and from the framework with the support of Table 2 also ensures that the study measures what it intends to measure. Managing generalization Managing generalization is important as it is the concluding “end productregardless of the language used to

describe it” (Polit and Beck 2010, 1452). They define generalization as “an act of reasoning that involves drawing broad conclusions from particular instances” (2010, 1451). Yet, it is more frequently discussed by quantitative researches than among qualitative ditto (Polit and Beck 2010). As Ritchie et al. (2014, 348) points out, the problem within qualitative research is that generalization is controversial and appears to be given diverse meaning and thereby inconsistently applied. However, as Polit and Beck states “leaders in qualitative research have begun to note the importance of addressing generalization, to ensure that insights from qualitative inquiry are recognized as important sources of evidence for practice.” (2010, 1451) To clarify matters, Polit and Beck argues that there are three different models of generalization: “the classic statistical generalization model, analytic generalization, and the case-to-case transfer model” (2010, 1451). Statistical

generalization is intimately related to selecting a sample “that is representative of the population” (2010, 1452), and where the strategy to achieve this is through random sampling. Analytical generalization refers to “rigorous inductive analysisaddress[ing] the credibility of the conclusions” (2010, 1453). The case-tocase transfer model is applicable if the researcher can “provide detailed descriptions that allow readers to make inferences about extrapolating the findings to other settings.” (2010, 1453) Out of the three different models Polit and Beck (2010) suggest, referencing to both quantitative and qualitative studies, the latter two are more applicable to the project than the former. Moreover, the second model (analytical) refers to credibility of the conclusions which relates to one of Robinson’s (2017) four requirements (‘valid,’ ‘credible,’ ‘feasible,’ and ‘utility.’) (Stewart Robinson 2017, 2744–45) of conceptual modelling. The third model

(case-to-case transfer) is intimately related to external validity, since it refers to the perceived utility (also one of Robinson’s requirements) of the research for a user or consumer of the research. Ritchie et al suggests that generalizations involves being able to make inferences from “robust and credible evidence” (2014, 349), which is established by the two concepts, validity and reliability previously addressed. To summarize, generalizations is an act of reasoning (Polit and Beck 2010), that involve EXPLORING DECISION ADVANT AGES | 125 inferences and drawing conclusions from the results of the experiments. The most applicable theoretical models of generalizations to reference in this context is the analytical and the case-to-case transfer model. These two relate to Robinson’s (Stewart Robinson 2017) requirements in modelling and simulation. Robust and credible evidence from the experiments provide the foundation for making inferences and drawing conclusions. The two

concepts of validity and reliability are therefore vital elements in making generalizations. Inferences are important as Ritchie et al point (Ritchie et al. 2014)out when discussing generalizations. The project employs measurements of performance (see section Experiments - a need for comparison?) as inferences in its evaluation of each of the experiments. The criteria to be met and measured are all to a certain degree related to effectiveness. For instance, they could involve the precision in which the model solves its task, the efficiency in which it perceives its inputs, processes it according to the instructions given and act by providing a solution, timeconsumption, and consistency. Obviously, using these performance metrics, experts in targeting could probably understand the effects the intervention has within the given problem setting. Others may view evidence such as the data in Table 2 and 3 in Chapter 5 as enough to draw their own conclusions. However, the intention in the

project is to infer based on the performance metrics which are rather straightforward. Plainly stated, if the validity and reliability of the experiments are sufficient and the AI-model’s performance produces measurable effects, then generalizations can be drawn as inferences. Operationalization This section engages in the creation of an analytic tool (operationalization) from which the subsequent empirics (Chapters 5 and 6) can be explained and understood. Operationalization can be defined as the intended employment of the theoretical framework, the unit of analysis, and the unit of observation. The section also covers a brief outline of the two simulation experiments. The OODA has been credited for its usage on all levels of warfare, as well as an approach to C2 (Osinga 2007). The joint targeting cycle (JTC) on the other hand represents a conception of a specific function within joint warfare. At a systems level, they both accomplishes a full cycle of decision-making, and can be

used as two separate frames for the empirics. The specific context for each experiment (usecase) is derived from challenges within joint targeting as illustrated in Figure 8 126 | EXPLORING DECISION ADVANTAGES Figure 8. Illustrating the role of the OODA and JTC respectively, and the relative levels of abstractions from experimentation to War Science. On the right-hand side in the model, the scope or reach of the experiments, doctrines and theory are defined. The results from the experiments have its main impact on decision-making within joint targeting. Along with the levels of abstraction via C2 towards joint warfare concepts the significance of the findings will decrease, although still indicative in certain aspects. Unit of Analysis (UoA) To demonstrate how two AI-models can augment human decision-making within joint targeting requires perceiving the unit of analysis as the OODA loop in concert with the JTC and the specific method for dynamic targeting (F2T2E2A). The argument

is that all three ‘cycles’ relate to the same overarching purposes of understanding a situation, deciding on how to deal with it, then act and assess. They share much in common and have been used in research in this field previously (Szeligowski 2018; B. Johnson et al 2023) Moreover, they implicitly embody how command and control (C2) is employed. Boyd acknowledged the OODA loop’s relations to C2 in his work, An Organic Design for Command and Control (Boyd 2018), as does the western doctrines on joint targeting (NSO 2021; US DoD Joint Publication 2018a). EXPLORING DECISION ADVANT AGES | 127 Below (Fig. 9), the three loops are placed side-by side to illustrate their commonalities as applied forms of command and control, implicitly mounted on three pillars: understanding, decision, and action. Figure 9. The three units of analysis They all share the same overarching purpose of understanding before acting. Importantly, the project utilizes all three. As described in Chapter 3

and visualized in Fig 8, Boyd’s strategic theory and his OODA loop has a reach beyond the scope of what JTC, including its subordinate dynamic loop has. However, they share the same purpose of understanding before acting. By gathering data and information a joint force can assemble various common tactical and operational pictures to describe the environment. The most important areas for a joint force’s intelligence and information collection is often referred to as the ‘Commander’s Critical Information Requirement’ (CCIR) in contemporary terms (Chairman of the Joint Chiefs of Staff 2021, 47). In a Boydian perspective this would refer to the context of orientation, because it “shapes observation, shapes decision, shapes action, and in turn is shaped by the feedbackcoming into our sensing or observing window” (Osinga 2007, 230). By processing, analyzing, and synthesizing the accessible data and information, a joint force can disseminate its common outlook and understanding

of the situation at hand, and what to do about it. On the other hand, maintaining the situational understanding is essential and requires constant updates. Otherwise it will limit the ability of a joint force to exploit opportunities, “magnify an adversary’s friction”, “stretch-out his time”, and “operate inside adversary’s OODA loop” (Osinga 2007, 196– 97). Hence Orient conditions C2 and impacts the effectiveness decision-making and of C2 at all subordinate levels. When reviewing NATO’s JTC and its dynamic targeting loop in the same way, the first three phases of each resemble Boyd’s two Os. Using the same analogy, Boyd’s Decide bears a resemblance to the JTC’s phase 4 and the part of phase 5 referred to as mission planning. In the case of the dynamic targeting cycle, this thesis suggests that the fourth phase Target would equal Boyd’s component Decide. JTC’s phase 4 plus phase 5 Mission planning, as well as Target in F2T2E2A, accounts for several

subprocesses. These include prioritizations of targets that are detected and identified, selecting the most feasible effectors and weapons, 128 | EXPLORING DECISION ADVANTAGES and considering available time, and constraints. The OODA loop’s Act equals the last part of phase 5 and phase 6 in the JTC and Engage/Assess in F2T2E2A. Essentially, Boyd’s feedback and feed-forward loops are assessments that reconnect and informs, in the same way the targeting cycles utilizes them. Notwithstanding this shared purpose, they arguably differ in the way they approach decision-making. The OODA-loop is more knowledge-centric whilst the JTC is decision-centric. Even if Boyd’s OODA-loop is used as the main framework for decision-making in the study, his theory incorporates gateways, or shortcuts, intentionally enabling a bypass of the Decide phase, given the right conditions. These conditions, as elucidated in Chapter 3, are defined by knowledge. This can originate from experience,

professional judgement of the situation at hand, or other forms of knowledge such as intuition. Contrary to this, the JTC is perceived as decisioncentric, specifically tailored to incorporate cautious targeting decisions, often via a targeting board and the explicit use of delegations of targeting authority, be it “target validation board” or “target engagement authority” (TEA) (NSO 2021, Edition B:LEX17-19). By reasons explained in this section and in Chapter 3, both the OODA loop and the JTC loop are important frameworks for the experimentation. Only employing the OODA loop also risks diluting it into a specific context to which there already exists a framework – the JTC. On the other hand, focusing on the JTC risks missing the opportunities inherently integrated in Boyd’s framework. In sum, the three loops share common ground but are used to expound and explicate different aspects of the empirical part. Where the JTC framework, and its more explicit dynamic method

(F2T2E2A), can explain part directly relating to targeting, Boyd’s OODA loop can justify other, broader aspect of decision-making within command and control more generically. Next, the two experiments are introduced along with an analytical tool explicitly built to support the understanding of how the two experiments interrelate to the theoretical framework. Options that make use-cases The doubts in the literature about the use, meaning and efficacy of AI as a tool for joint targeting purposes makes simulations a fertile ground to show how AItechnology could augment the joint targeting process. Complete solutions are still distant, despite current endeavors (Department of Defense 2022; Penney 2023; Hammes 2021), and novel applications such as the Israeli’s ‘Gospel’ and ‘Lavender’ (Davies, Mckernan, and Sabbagh 2023; McKernan and Davies 2024). Theoretically, AI-solutions within targeting could support variations in tempo, to include acceleration when opportunities arise,

without a decline in precision or quality of decisions. Recent research (B Johnson et al 2023) on where AI-methods could be of use is promising. It suggests 28 specific functions that show potential for enhancing and enabling the future naval kill chain(B. Johnson et al 2023, 11) The analytics are EXPLORING DECISION ADVANT AGES | 129 consistent with this project’s approach in the use of AI and ML for augmenting human decision-making in a dynamic targeting setting. By traversing their general results to explore two specific scenarios as the units of observation it is possible to address how AI and ML can be applied. The two experiments are briefly introduced below: Experiment 1: Enhancing precision and efficiency in a Joint Force Dynamic sensor allocation and target engagement with deep learning (Model 1) The first use-case/scenario, referred to as Experiment 1 with its solution correspondingly named Model 1, is the problem of finding high-value targets once they have been

identified in the target system analysis 35. This use-case primarily relates to the intelligence support activities to targeting, in which AI also would augment the collections operations management (COM) including parts of the tasking, collection, processing, exploitation and dissemination (TCPED) associated to joint intelligence, surveillance and reconnaissance-process (NATO 2020; US DoD Dictionary, 2020). The problem ties into the effective allocation of sensors, processing, and exploitation of data, as well as the more time-consuming collection, analyses and syntheses required to produce the target system analysis (TSA), or what Boyd and Sun Tzu refers to as estimates and foreknowledge. The model has its context (or design) framed by phase five of the JTC, and the dynamic targeting method (F2T2EA). The problem is that of locating a target that is considered as a high value target (HVT), which makes it a concern and priority for the joint force commander (JFC), though the location

is unknown. Allocating valuable and expensive intelligence collection assets (sensors) to search for the target will not only expose them but means that the asset is engaged in a low probability search. Balancing the resources needed is important for a joint force The solution is intended to be an AI-model that can augment rapid decision-making on sensor allocation with more precision and a high probability. In such a case, and if the JFC has authorized the use of the model’s output as input for a strike package, then that would become a decision policy, preformulated to enable an accelerated dynamic targeting method. Furthermore, if the AI-model is used prior to an operation it can be utilized for decision-making within wargaming as well as provide valuable insights to estimations and foreknowledge such as a target system analysis (TSA). If, on the other hand, the model is to augment decision-making during phase 2, Target development and Prioritization, then it should be able to

augment that as 35 No entity will ever become a target without the proper intelligence verifying its function for the adversary, its composition, and characteristics. 130 | EXPLORING DECISION ADVANTAGES well. The data needed for this experiment, the specific AI-method used to solve it, and other details are provided in Chapter 5. Experiment 2: Optimizing a Joint Force Dynamic Targeting Decisions with machine reasoning (Model 2) The second use-case/scenario, referred to as Experiment 2 with its solution correspondingly named Model 2, is the problem of selecting and prioritizing targets and matching the appropriate response. It is a classical problem termed weapon-totarget-assignment (WTA) problem, with roots in operations research from the 1950s and onwards, aiming to optimize the resources used, given specific constraints and conditions. Considering the ambition of joint targeting to match and integrate the appropriate joint fires capabilities from two or more components in

coordinated action, any WTA-problem becomes more complex when put in a joint context. The model has its context framed by phase five within the JTC, Mission Planning and Force Execution, and by residing within dynamic targeting. As alluded, the assumption made in this project is that a dynamic approach to targeting should be the preferred option and major tactic. The problem is based on inputs to the model Given a selection of weapons at the disposal, their respective characteristics and limitations, and some regulatory targeting policies for their use, a set of targets appear int the battlespace. Each target has their respective characteristics including protection, moving or stationary, geographically distributed to be within or outside the range of ‘our’ various weapons. What is the optimal solution on how to assign weapons to targets? It is a classical WTA problem in OR research, of which there are several suggested models within the literature of operations research. None of

these takes into consideration the more complex, yet realistic approach that this experiment does. This experiment is perhaps even more concerned with balancing the resources at hand within a joint force. Economy of force is being displayed and accounted for, which means that each decision will come with a price tag. The solution sought for is an AI-model that can augment decision-making at a fast pace, whilst keeping track of all available weapons that could potentially be used for prosecuting the targets that emerge. As with Experiment 1, the data needed for this experiment, the specific AI-method used to solve it, and other details are provided in Chapter 6. EXPLORING DECISION ADVANT AGES | 131 Unit of Observation (UoO) The data collection is done from the analysis of the two experiments. Each model solves problems derived from the UoA and is identified as pertinent and feasible for the project. The two simulation experiments provide the basis for managing the unit of

observation. The operationalization - the employment of the theoretical framework is visualized in Table 2. Each column represents, from left to right, the four stages of the OODA loop, the corresponding six phases of the JTC, the dynamic targeting method with its suggested 33 subtasks (columns 3-5), and lastly, the 20 intelligence support activities to joint targeting (columns 6-7). To exemplify the interrelations between the unit of analysis (UoA) and the unit of observation (UoO), we can use the following example: Boyd’s Observe is interrelated to JTC Phase 1, to the Find and Fix of the dynamic targeting method with its first nine subtasks, and to the first seven of the intelligence support activities. In parts of this consolidated table, there are a few subtasks that ‘fall in-between’ phases. This is to be expected when consolidating several different concepts and methods. However, this does not affect the aim and purpose of the analytical tool, which is to provide a common

reference point for the empirical study, the results, and subsequent conclusions and discussions on the study as a whole. 132 | EXPLORING DECISION ADVANTAGES Table 2 The interrelation between the unit of analysis (UoA) and the unit of observation (UoO) found in the suggested derivation of 33 subtasks within the dynamic targeting process and the 20 subtasks within Intel support to targeting. OODA Observe JTC Phase 1 Dyn TGT Find Cmr’s Intent, Fix Phase 2 Orient Phase 3 Capability analysis Track Cmr’s decision, Mission planning # 1 Search Indicators & warnings 1 3 Identify MoP/ MoE 3 Target Material Production 5 2 4 6 Target Detect JIPOE Determine TGT characterization Classify Confirm TGT validation Significance 9 TGT mensuration 8 Target System Analysis Basic target development 7 2 4 6 PID Estimations 7 10 Prioritization Intermediate target development 8 12 Estimate TGT window of vulnerability 14 Desired effects 11 15

Force assignments Phase 5 Intel support activities 13 Phase 4 Decide Subtasks/Functions 5 Objectives, and guidance Target development # Monitor ISR management Functional Risk estimation Time management 16 RoE/TGT Policies 17 Collateral damage Estimation (CDE) 19 Options 18 20 Threat analysis characterization 9 10 Damage estimation 11 Functional analysis including Advanced target development 12 Threat analysis 13 Recommendations EXPLORING DECISION ADVANT AGES | 133 21 Requirements incl Combat assessment 23 Select option 22 24 25 Force execution Engage Exploit Act Assessment Assess 26 27 28 Intel requirements 14 Collection tasking 16 Combat Assessment Battle damage assessment (BDA 1) 17 Re-attack recommendations Re-attack recommendations 19 Deconfliction/Coord/ Synchronization CID Field CDE Target engagement Measurements & Documentation Follow-on actions 30 Re-attack 32 MEA 33 15 Issue order 29 31 Sensor

allocation BDA 2 BDA 3 18 20 Further explanation to Table 2 Some complementary explanation to Table 2 is warranted to support better understanding of the defined subtasks (column 5) and intelligence support activities (column 7). Starting with the 33 subtasks, these are derived from explicit and implicit doctrinal guidance discussed in Chapter 3. The project derived in total 33 subtasks that in a more granular fashion declares what is otherwise incorporated in each of the seven elements of F2T2E2A. The granularity supports a more precise explanation of what the problem is and how it is solved by AI in the second experiment. In the first element FIND, five different subtasks are suggested to be present. Search, Detect, Identify, Determine Target characteristics and Classify all relate to tasks that collectively supports a confirmation of what the joint force’s sensors operating has found. Even if they explicitly search for a specific target, the distinction between detecting for

instance electro-magnetic rays or acoustic sound and identifying a radar station or a vehicle are not the same. Determining the targets’ 134 | EXPLORING DECISION ADVANTAGES characteristics could mean declaring if it is a moving or stationary target, or if it is protected, camouflaged or other important aspects. The classification is yet another issue that supports the understanding of for instance the specific type or model of vehicle, which in turn will become important factors to decide upon what means (effectors) to use. The next element is FIX, where four subtasks are defined. The first subtask (6) is a validation of the target, meaning is it a valid target? If not, it cannot be engaged for targeting purposes. If validated as a target, its significance becomes an important parameter, as well as positive identification (PID) of it which can be argued to coincide with the target validation but is kept separate here as PID has a legal perspective to it. Lastly, target

mensuration is required, meaning the exact geographical positioning of the target (the coordinates) is established. This enabled us to proceed to TRACK, given that the target proposes such an action. The first subtask is here to prioritize if the target should be engaged. If so, then resources need to be tasked to monitor the target. Moreover, an estimation of when it is vulnerable to be attacked (12) often has to be defined, along with a risk estimation (13). This leads to the fourth element TARGET, which is suggested to have 11 subtasks associated with it. These subtasks are defined as confirming the desired effect (there is a difference between for instance destroying and delaying a target), managing time (15) of own potential resources (also relates to threat analysis and the targets vulnerability aspects), ROE/ targeting policies that needs to be addressed and complied with (16), collateral damage estimations (CDE), threat analysis, and the options (19) that are available for the

decision-maker. This will result in targeting recommendations (20), along with requirements to include the way combat assessment (CA) is managed. Concerns such as deconfliction of forces, coordination issues and synchronization of targeting activities in time needs to be addressed and solved prior to one of the recommendations is selected as the option and an order is issued (24). The order means we move into the fifth element ENGAGE, necessitating a second identification of the target (CID) by the resource used for attacking the target to assert and verify that it is the correct target. A second CDE is also done to verify the conditions and estimate any collateral damage prior to the target engagement (it could in theory be hours between the first CDE and the second, opening up for changed conditions). Thereafter the target is engaged (27) The sixth element EXPLOIT involves considerations on how to measure what during the target engagement. This is why it is in-between ENGAGE and

EXPLOIT, as this should be addressed prior to. However, the actions taken to document a target engagement for instance by taking photos or videos of the event to support evaluation and assessments post-strike are done here, as well as any other form of follow-on action to exploit the target or the situation. The last element ASSESS also has follow on actions (29) included, mainly to allude that, pending the specific target engagement, an initial assessment could invoke exploiting. Nevertheless, the most significant subtasks are reattack (if there is a need to reattack for instance if the attack was EXPLORING DECISION ADVANT AGES | 135 unsuccessful by missing the target), and combat assessment (CA) (which involves several subcomponents, some of which are found in the intelligence support activities). Munition effectiveness assessment (MEA) is doctrinally incorporated in CA. However, it is made explicit here to differentiate it so as to enable more subtasks to be considered as

problems that AI could be engaged in. The last subtask is reattack recommendation, which often occurs as a consequence of battle damage assessment reporting (BDAs). Similarly, the first experiment is supported by the 20 intelligence support activities derived from the same seven elements of F2T2E2A. In the first two elements FIND and FIX, seven different activities are suggested to be present. These are also part of what constitutes the foundation for Phase 1 in the JTC. Nevertheless, these initial activities (and products in some cases) are of continuous value as an operation moves from planning into execution. Indicators and warnings (1) are perhaps more obvious to a nonexpert than the other four activities. The joint Intelligence preparation of the battlespace environment is an all-encompassing estimation serving as a foundation from which more detailed analyses, to include target system analysis (TSA), are conducted för specific functional areas of the adversary’s system.

Various target material productions (TMP) are produced in response to specific intelligence requirements. One of the most fundamental TMPs is the basic target development (BTD) made for each individual target (see Chapter 3 for more details on basic, intermediate, and advanced target development). Apart from these different criteria to support evaluations are defined, for instance measurements of performance (MoP) and effectiveness (MoE), along with different estimations predicting actions, intention, or whereabouts of the adversary (foreknowledge). TRACK involves intermediate target development (ITD), managing the ISRresources (9), functional characterization (10) which is a complex endeavor of detailed descriptions of the function of each target. A precise description will support the decision on, for instance, what weapons to use, how much munition, and where to hit. Damage estimation has been included in this element as well In the fourth element TARGET, functional analysis (an

analysis over to what degree, if the target is affected, this affects the larger system. It supports prioritizations), including advanced target development (ATD, see Chapter 3) are being done. Threat analysis (13) along with the estimated intelligence requirements are declared. This is significant for the decisions on what, where and why sensors should be allocated (15). The last three elements ENGAGE, EXPLOIT and ASSESS are relatively straight forward. Collection tasking (16) includes all orders given to intelligence collection resources (sensors) and is present throughout the three elements. The content of the three battle damage assessments (17, 18 and 20) has been discussed in Chapter 3. As a last remark, reattack recommendations (19), is defined here as the intelligence portion to such a recommendation. Whilst it can be argued from an operations perspective that a reattack is unwarranted, the intelligence perspective may state the opposite but for other reasons. 136 | EXPLORING

DECISION ADVANTAGES Given the abovementioned remarks, Experiment 1/Model 1 is purposely focusing on the intelligence support activities within joint targeting. Its reference is column 7, named Intel support activities with a total number of 20 potential subtasks, where AI/ML could be applied. Note that these are all of the activities defined by this project when deriving activities from the elements of F2T2E2A. Model 1 is intended to augment and optimize the planning and execution of sensor allocation for target intelligence collection, especially the estimate and location of an adversary’s more valuable assets. Therefore, it also relates to an augmentation of the target material production such as Target System Analysis (TSA) during JTC’s phase 2. Moreover, it could be perceived as having a significant role during JTC phase 5, and correspondingly within the part of dynamic targeting concerned with finding and fixing a target. This means that Table 2 enables us to assess how

the two models concurrently augments the other cycles, or loops. Similarly, the output of Model 2 is tailored to augment the dynamic targeting method to enable both faster prosecutions and support resourceful allocations of resources. As such it also relates to Boyd’s Observation, Orientation, and Decision, and serves part of JTC’s phase 3,4, and 5. The suggested derivation of 33 subtasks within the dynamic targeting method is intended to enable a more thorough analysis of the results. Reflections Reflecting upon the design of the project and the methods used yields an opportunity to broaden the perspective and address concerns within and beyond the scope of this thesis. The relevance of what is covered in this section is for the reader to consider. It has been a reservoir for thoughts throughout the project A changing paradigm of scientific research? Using artificial intelligence to augment, or even conduct research is becoming a reality that may change the paradigm of

scientific research. It is already an emergent research method where key work is being undertaken (NATO, 2020). However, that was back in 2020, and at the time this project commenced. Much has changed in such a brief period as five years, partly due to the development within the crossdisciplinary field of AI. The birth of the now renown ChatGPT on the 30th of November 2022, a large language model (LLM), made scholars cautious, lecturers weary, and students eager. Now, anyone with a smartphone has a virtual assistant to ask whatever and whenever. More importantly, AI is being deployed and used in most fields research. In 2023, one of the first methodological books on AI research in social science was released (Green and White 2023). More is likely to follow, given EXPLORING DECISION ADVANT AGES | 137 the development in other fields such as medicine research, where AI-discovered drugs have shown success rates between 80-90%, significantly higher than the historical industry averages

of 40-65% 36. Recently, engineers at MIT have designed micro-batteries, with the use of AI, which could enable the deployment of cell-sized, autonomous robots for drug delivery within in the human body 37. Arguably, these are signs of a shift in scientific research, including social science, where AI assists, or become a main resource in the scientific context of discovery as well as in the following context of justification (Swedberg 2012). AI is in such a milieu, in all essence, augmenting the human researcher. However, as Green and White shows, AI (in their case machine learning) is becoming a research method in itself, effectively giving rise to novel social science methods (2023). This will create new opportunities in war studies, where applied AI would be capable of categorizing data in original ways, cast new lights on causalities, and make inferences or discover interrelations that has previously been beyond human limits. Given the opportunities laying ahead of using AI for

research and as a method, a relevant question to ask is how the results can be trusted. Trusting AI has been, and still is a vital part of what some researchers are wrestling with (Ayoub and Payne 2016; Svenmarck et al. 2018a; Gill 2020) The constant evolution of science and warfare It is important to bear in mind that science, just like warfare, evolves and changes with time. The sophistication of contemporary AI-applications, and the current frontier of context-aware artificial agents challenges the human monopoly as decision-makers. It transforms the perception of agency Occasionally, development leads to what Kuhn, one of the most influential philosophers of science of the twentieth century, suggests being scientific revolutions, arguing that these are: [inaugurated by a growing sense, again often restricted to a narrow subdivision of the scientific community, that an existing paradigm has ceased to function adequately in the exploration of an aspect of nature to which that

paradigm itself had previously led the way] (T. Kuhn 2020, Foundation:92) What Kuhn suggests is that development, even in another field of science, may lead to radical changes, or revolutions in an existing idea or concept. His engagement in the history of science, and his affection to the work of Aristoteles, was in itself a rather radical shift, as his doctorate in 1949 concerned an “application of quantum 36 Brian Buntz, “6 signs AI momentum in drug discovery is building”, Drug Discovery & Development, 24 June 2024, dihttps://www.drugdiscoverytrendscom/six-signs-ai-driven-drug-discovery-trends-pharma-industry/ 37 Science Daily, “Engineers design tiny batteries for powering cell-sized robots” https://www.sciencedailycom/releases/2024/08/240816121453htm, accessed 30 August 2024 138 | EXPLORING DECISION ADVANTAGES mechanics to solid state physic” (Bird 2022, 2) 38. The point made here is that the development of AI as a technology represents a potential revolution

in parts of modern warfare. The project explicates military decision-making and argues that it represents what Kuhn refers to as a subdivision that is, at least challenges, or even on the verge of adequately characterizing the agency in decision-making. There is a current tension to the human-centric approach of C2 when comparing it with the requirement of AI brought forward in strategic visions. Additionally, the limitations of human cognition related to comprehension and tempo (or speed), is in itself a source for change and evolving conception. It brings about thoughts on new ways to collaborate with increasingly intelligent machines. The cognitive limitations of humans, the growing amount of data and accelerated warfare at the tactical edge, and the increased integration of AI as mitigations for mismatches, efficiency or shortcomings raises questions of agency in decisionmaking. If AI is perceived as an actor, it will become a property member of the decision calculus. This will

trigger questions related to who is commanding and controlling what parts of an operation. If, for instance a tactical command is utilizing large language models (LLMs) to develop four options to approach a tactical situation, and then a human commander decides on one of those, is that then considered as a responsible action – one that the human commander can be accountable for? If so, was the human commander demonstrating meaningful human control? If another AI decision support system gives its recommendation out of the four options, and the human commander then agrees to that recommendation and then decides, how can we ensure human judgement, unsolicited biases and avoid over trusting the machines? The current tension and the subsequent question of agency also co-creates an asymmetry to the fundamental perceptions of artificial intelligence. AI is either seen as an actor that can do things or perceived as yet another tool in the technological toolbox that we (humans) can use when

needed. From the latter perspective AI is used by humans for humans, hence perceiving the material world as exogenous to society and instrumental to human intentionality. The former incorporates a more dynamic or fluid agency of both humans and intelligent non-humans that progressively imposes resurrection to the question of agency. Law (1992), coming from a constructivist approach, refers to technology within social practice as “They shape it.” (1992, 382), meaning that technology shapes humans as much as humans shape technology. This symbiotic relationship, and the potential effects it could have, made DeLanda (1991) suggest that this was a critical first step of an apparent “migration of control from humans to machines” in order to rationalize the “division of labor”. (DeLanda 1991, 155) 38 Stanford Encyclopedia of Philosophy Archive (Spring 2022 Edition), “Thomas Kuhn”, https://plato.stanfordedu/archives/spr2022/entries/thomas-kuhn/, accessed 18 August 2024

EXPLORING DECISION ADVANT AGES | 139 Technological advancements are constant companions in the evolution of human traits. Advancements can be progressive or radical They can origin from trial and error, fortunate secondary outcomes, or plain creativity. However, most advancements are hard to predict. If ‘predicting the future’ is difficult (van Creveld 2020, 232), then the appearance of technological innovations may attract early adopters and visionaries that ignites evolutionary changes from existing paradigms, as Kuhn (2020) suggests. In the book ‘Technology and War’(1989), van Creveld suggests that war and warfare is governed by technology. Notions such as van Creveld’s (1989), or Law’s (1992) on technology’s shaping effect, or Stiegler’s argument on new technological inventions, claiming that these “are born with the appearance of the limits of the preceding systems, owing to which progress is essentially discontinuous.” (1998, 33) These are perspective on

the role technology has had in the development and transformation of science and practice. However, war and warfare also play a role in the development and transformation of technology. Visions and strategies such as the US JADC2 (Department of Defense 2022), or China’s ambitions (E. Kania 2019) also shapes technological investments. The visions, strategies and ambitions indicate that a range of innovative military applications are expected to be liberated through the integration of artificial intelligence and machine learning. It stimulates novel concepts and challenges prevailing conceptions of war and warfare. As revealed in 2021 during a US joint hearing; “For better insights, intelligence agencies will need to develop innovative approaches to human-machine teaming that use AI to augment human judgment” (Testimony on Final Report of the National Security Commission on Artificial Intelligence, 2021, p. 4) It reiterates what Nurkin and Siegel (2023) argues when claiming the

need to prioritize “the ethical and responsible development and use of trustworthy AI and keeping humansand human judgmentat the center of human-machine teams (2023, 15). Meanwhile, the ongoing efforts to define meaningful human control in the midst of novel warfare concepts integrating AI is continuing (Bode and Huelss 2024). Conclusions The chapter covered the design and methods used when investigating a phenomenon amid changes, innovations, and uncertainties. The project warrants a ‘pragmatic approach’ to produce useful knowledge (Friedrichs and Kratochwil 2009). The exploring nature of this research makes modelling and simulation (M&S) an effective method, along with complementary methods to include interviews with SMEs to understand the real problem, and surveys external validation of the results. This multi-method design corresponds to the authoritative study on military power by Stephen Biddle (2010), which also anchors the project within War Studies. 140 | EXPLORING

DECISION ADVANTAGES The absence of specific guidance on experimentation using Modelling and Simulation (M&S) necessitated that the project primarily drew upon frameworks established within the field of Operations Research (OR). Given the lack of evident methodological conflicts between OR and the social sciences regarding experimental approaches, the project adopted OR-based guidance as its methodological foundation. Beyond the design and method discussion, the chapter also presented its analytical tool - an operationalization of the theoretical framework serving as an intermediary instrument to, and from, both experiments. Recent research exploring the use of different AI-methods to enhance the future ‘naval kill chain’ (B. Johnson et al 2023, 11) indicate where AI and ML could augment human decision-making in a dynamic targeting setting. Their general findings along with the decision of what use-cases (problems) and specific AI techniques (solutions) to explore using two

specific targeting scenarios as the units of observation, makes it is possible to address how AI and ML can be applied more explicitly. The two experiments were briefly introduced in this chapter and will now be presented in detail, each in its own dedicated chapter. Lastly, the reflections covered the potential changing paradigm of scientific research and the constant evolution of science and warfare. The transformative nature of warfare underpins how human and material resources are combined, organized, and deployed to achieve strategic objectives. Echoing Kuhn’s (2020) view on paradigm shifts, the emergence of new technologies can catalyze evolutionary changes within existing paradigms, attracting early adopters and visionaries who drive transformative progress. This unpredictable evolution underscores the necessity for both flexible frameworks and a continuous reassessment of the ethical boundaries and professional responsibilities associated with new technological possibilities

in military contexts. Over time, we may observe a shift in both command and control from human agents to increasingly autonomous and intelligent machines. Such dynamics necessitate careful ethical deliberation and professional judgment, as the implications of this interdependence extend beyond technical considerations. EXPLORING DECISION ADVANT AGES | 141 142 | EXPLORING DECISION ADVANTAGES Chapter 5 – Experiment # 1 Model 1 - Enhancing precision and efficiency in a Joint Force Dynamic sensor allocation and target engagement with deep learning Man [.]is the most imitative (mimetikotaton) of all animals and he learns his first lessons through mimicry (dia mimesos). Aristotle, Poetics 39 Introduction The chapter covers an experiment where an AI-model is given the role to enhance precision and efficiency in a joint force dynamic sensor allocation and target engagement. The experiment explicitly explores the use of artificial intelligence for enhancing intelligence activities

within the joint targeting process using a method termed deep learning. Deep learning (DL) is a subset of machine learning (ML) and is one of the more vivid AI-techniques, mainly because of its ability to improve through iterations of its own learning. Much like Aristotele’s reflection on the (hu)man way of learning through imitation and mimicry of those who have already acquired knowledge, machines in this context 40 replicate this approach - repetitions make for better performance. The more training a modern neural network is given the better its results. In this experiment it predicts the position of specific target engagement radars (TERs), which is considered to be a critical part of the targeting process. The chapter begins with a general description of deep learning as a concept and AI technique. It then discusses the military problem to be solved, after which it continues with describing the experimental set up along with the computational implementation and solution of the

problem. The problem concerns the 39 Citation in Bourdieu’s ‘The Logic of Practice’, Stanford University Press, 1992, 25. The quote draws a parallel between human learning through imitation and the way machine learning algorithms often learn patterns from data by modelling or mimicking input-output relationships. Aristotle’s insight about imitation as a foundational aspect of learning can therefore be seen as precursor to supervised learning in artificial intelligence. 40 Sarker suggests a similar analogy, when referring to DL as an “AI function that mimics the human brain’s processing of data”. See Sarker ‘Deep Learning: A Comprehensive Overview on Techniques, Taxonomy, Applications and Research Directions.’, Springer Nature: Computer Science, 2021, 2 (6): 1. EXPLORING DECISION ADVANT AGES | 143 geographical positioning of a specific target 41. Put simply, if a target is known to be within a joint force’s area of operations, the succeeding logical question

would be: where is it? If the geographical location of the target is unknown, or not precise enough, this will generate intelligence collection requirements to try to locate and identify the target’s exact location. The significance of the target will constitute a joint force’s efforts to find it, which in turn means that sensors will have to be allocated and assigned to search and collect that information. Given the type of target and the area of search, this may be difficult, time-consuming, and resource absorbing. Therefore, the rationale for this experiment is: if a deep learning model could define explicit locations (preferred by a specific weapon system, given defined technical and tactical conditions for that weapon system within a given geographical area), then intelligence collection would become more precise, and resources be used more efficiently. Moreover, corresponding intelligence reports, including opensource intelligence (OSINT) and crowdsourced social media

reporting, could be used as an overlay for increased association of intelligence interest, and greater reliability to the information given by the deep learning model. The solution is used in different simulation activities to analyze its performance. This brings the chapter to the results and subsequent evaluation. Lastly, it concludes the implications of the experiment by discussing the results and circles back to the aim and objectives of this experiment. Deep Learning (DL) Computational approaches to learn and improve performance over time Computational approaches in which machines learn refer to a broad set of methods that enable computers to learn. The focus is on developing algorithms and models to process data, identify patterns and improve performance over time, in order to make predictions or decisions without being explicitly programmed. The main types of computational approaches include supervised learning - where machines learn from labeled data by mapping inputs to

known outputs, unsupervised learning where machines identify structures and patterns from unlabeled data, and reinforcement learning - where machines learn through trial and error using reward or penalty functions. The taxonomy adapts to the development in AI which we will come back to. Nonetheless, these different computational approaches have become an indispensable instrument in several industries, and can be utilized for object 41 A geographical positioning is relevant to all target types, with one exception – virtual targets. 144 | EXPLORING DECISION ADVANTAGES detection, picture, pattern recognition, data collection, data sorting, and audio-totext translation (Taye 2023, 1). Deep learning (DL) uses artificial ‘neural’ networks inspired by the structure and function of the human brain described in neuroscience (Goodfellow, Bengio, and Courville 2016, 164). Neural networks consist of multiple layers of nodes (or neurons) that progressively transform input data to

produce the network’s output (Ibid 2016, 164). Each layer extracts various levels of abstraction from the data: the initial layers capture low-level features (such as edges or textures in images), while deeper layers combine these into more complex patterns. Rather than representing degrees of knowledge or certainty, each layer performs specialized transformations that enable the network to identify increasingly abstract representations in the data. The final layers use these patterns to, for instance, classify an object or make predictions of something. These learning algorithms are capable of learning hierarchical representations of data, making them effective for tasks such as image recognition, natural language processing, and speech recognition. The advent of deep learning came in 2006, when Hinton, a pioneer in neural computation, presented his work on ‘deep belief nets’ (Hinton, Osindero, and Teh 2006, 1527). It marked a major step in computer science with extensive

implications. As Sarker (2021) suggests in an authoritative overview of the topic, that deep learning is a core technology in the current Fourth Industrial Revolution (4IR or Industry 4.0), widely applied in various application 42 areas like healthcare, visual recognition, text analytics, cybersecurity, and recommendation systems (Sarker 2021, 12). More importantly, Sarker argues that deep learning is ‘a frontier’ for artificial intelligence and pushes the technology to a new, smarter level, in which deep learning methods can “play a key role for advanced analytics and intelligent decision-making.”(2021, 3) An illustration of the performance comparison between deep learning (DL) and other machine learning (ML) algorithms can be seen in Figure 10, showing that DL modelling from large amounts of data can increase the performance, compared to ML. (Sarker 2021, 4) 42 In the overview Sarker compiles a summary of deep learning tasks and methods in approximately fifty different

popular real- world applications areas. See Sarker ‘Deep Learning: A Comprehensive Overview on Techniques, Taxonomy, Applications and Research Directions.’, Springer Nature: Computer Science 2021, 2 (6): 12 EXPLORING DECISION ADVANT AGES | 145 Figure 10: A typical performance comparison between deep learning and machine learning considering the amount of data. Source: A Comprehensive Overview on Techniques, Taxonomy, Applications and Research Directions (Springer, Singapore: Sarker, 2021), p.4 Sarker (2021)provides an aggregated taxonomy on DL techniques based on how they are used to solve problems, defining these as “(i) Supervised: a task-driven approach that use labeled training data, (ii) Unsupervised: a data-driven process that analyzes unlabeled datasets, (iii) Semi-supervised: a hybridization of both the supervised and unsupervised methods, and (iv) Reinforcement: an environment driven approach” (2021, 5). Whilst all four approaches in Sarker’s (2021) DL-taxonomy

are various ways machines learn; the former (supervised) learns its ‘first lessons’ by being taught by someone or something in how it should interpret and understand the specific task (imitate/mimic). This supervised learning approach is also the technique used for the DL-model in this experiment. It learns by being given a substantial number of right answers from the precursor model’s statistical program, to be able to perform on its own when the training is over. The military problem Modern military targeting is planned for, executed, and assessed within the larger frame of military operations. Whether joint (more than one service engaged) or not, the overall objectives and intent from the military commander guides and directs all succeeding targeting work and efforts during an operation (see Chapter 3, ‘Joint Targeting’, for a more complete presentation of the process). One of the many concerns within the joint targeting process is knowing the location of the selected and

prioritized target. Without the information of the whereabouts of a specific target, the need for sensor resources increases. The use of intelligence collection assets to locate prioritized targets is a key to targeting successes. Nevertheless, target intelligence collection assets are often expensive and limited, making opportunistic random searches inefficient and potentially a waste of time and 146 | EXPLORING DECISION ADVANT AGES valuable assets. As such, the problem of not knowing the location of a target is the focus of this experiment. The problem is delimited by defining the type of target to medium-range groundbased surface-to-air missile systems (MSAMS). MSAMS targets of this type can be narrowed down to their critical elements 43, or the essential parts that enable them to perform their primary function. The critical element of MSAMS is its radars, which create situational awareness, detect and identify incoming threats, and enable target engagement by providing the

necessary input data to the missiles. By consulting military experts 44, the problem is further advanced by splitting it up into three parts and, each of which to various degrees can narrow the location of the MSAMS. The three parts are: 1) generic information of the tactics, techniques, and procedures (TTPs) of MSAMS; 2) the specific information of the TTPs for an explicit type of MSAMS; and 3) the technical data based on the specific terrain topography that dictates the conditions of radiofrequency (RF) propagations from MSAMSs radars. Based on this division, the approach to the latter part of the problem was decided by using neural networks based on the assumption that the DL-technique of semantic segmentation has the ability to segment terrain topography into different regions based on their semantic meaning. Current methods to model the behavior of electromagnetic radiation (including radio emitters) from a point of transmission over irregular terrain relies on mapping out a

region with practical measurements 45 or relying on simulations using various statistical programs. The Longley-Rice model 46 , is an RF propagation model for predicting the attenuation of radio signals for a telecommunication link. It is one of the most widely used models (Kasampalis et al. 2013) Other popular choices include the Irregular Terrain Model and the Irregular Terrain with Obstructions Model (Ibid. 2013). Although practical measurements and these models can provide a basis to find optimal locations for radio emitters, the time it takes can be too long if real-time decisions are needed. There is a need for more efficient methods of radio emitter location determination. 43 The elements of a target’s composition that are vital for its function. See Chapter 3 44 Subject matter experts at the Swedish Air-Defence Regiment LV6 in Halmstad. 45 For instance, using traditional tools such as clinometers or military protractors provide foundational methods for assessing LoS in

fields. 46 The Longley-Rice model is designed for frequencies from 20 MHz to 20 GHz, distances from 1 to 2000 km, and antenna heights from 0.5 to 3000 m It provides path loss estimates by combining physics with empirical data Source: MathsWorks, “LongleyRice”, https://www.mathworkscom/help/antenna/ref/rfproplongleyricehtml, accessed 24 May 2024 EXPLORING DECISION ADVANT AGES | 147 The real problem of intelligence collection for targeting purposes The problem of managing intelligence collection for targeting purposes is multifaceted. Based on the operational plan, the forecasting from the assumptions and assessments made by a joint force, the Joint Forces Commander (JFC) issues tailored targeting objectives, direction, and guidance. This can be perceived as shaping the approach to targeting (or part of the Orientation phase directing the Observations in the OODA loop (see Fig.1, Chapter 3)) This initiates the joint targeting cycle (see Chapter 3). The staff and subordinate

components engaged in, or supportive to, the targeting process define, develop, and prioritize targets that, if targeted, would create desired effects in accordance with a JFC’s objectives. The designated targets will eventually constitute a joint integrated prioritized target list (JIPTL) for a finite episode of an operation. The cycle is repeated throughout an operation as it progresses, giving rise to new situations, new objectives and, consequently, new targets. However, the intelligence on each of the prioritized targets will vary Targets that are fully developed could be scheduled to be targeted using the deliberate targeting method. Other targets will require intelligence collection efforts prior to any target engagement. A fundamental criterion in all target development is the specific location of a prioritized target. The target needs to be found In this case the collection efforts are defined as sensors that need to be allocated to meet that criterion. The relationship

between the model and reality is critical to its utility A joint force will largely have a limited number of collection resources at its disposal. Given that a certain percentage will be allocated for other types of intelligence collection tasks, only parts of the total number can be assigned to support a dynamic targeting method’s 47 first two phases: find and fix (locate and identify). Employing an AI model as a targeting advisor can significantly enhance the precision and efficiency of dynamic sensor allocation and target engagement within a joint force operation. The potential utility is described using a simple example (Fig 11). In this scenario, the AI model provides decision-makers with informed recommendations on where to focus search efforts, specifically for identifying MSAMS Target Engagement Radar (TER) sites. The geographical area shown in Figure 11 is believed to harbor MSAM threats. Intelligence suggests that three potential locations (large orange circles), align with

the current assessment of the adversary’s intentions. Using deep learning analysis of satellite imagery, the AI model identifies five recommended locations (small red circles) for potential TER sites. Notably, two of these recommendations 47 Though the dynamic method is defined as a method within the joint targeting process, the method does resemble a process. 148 | EXPLORING DECISION ADVANTAGES overlap with the suspected locations (orange circles), reinforcing existing intelligence assessments. Another location stemming from an open-source intelligence (OSINT) report is an indicator of the presence of alleged large rocket launchers (green circle). This correlates to one of the AI-models‘ recommended locations. The final two AI predicted MSAMS locations, although isolated, provide points of interest. When time is critical, the AI model’s insights can guide prioritization, helping determine which areas should be assigned sensors for intelligence collection tasks to support

dynamic targeting effectively. Figure 11: A basic example of how the model could augment decision-making during intelligence support to dynamic targeting. The intended solution is to have a model predict (data-driven insights) the most likely locations of a MSAMS’s TERs in a relatively short period of time. This makes sensor allocations more precise and efficient during dynamic targeting and enables a faster target engagement. More specifically, the model computes and presents recommendations on where to allocate sensors, and by this acting as a targeting advisor expounding likely TER sites for a decision-maker. Furthermore, the solution is intended to make use of neural networks to facilitate a reduction in time, as well as the potential for scalability where a model can include even more adversarial systems. It renders the model both a tool for ‘estimate’ or ‘foreknowledge’ (to use Sun Zhu and Boyd’s terms), and for targeting services. As Robinson (2017) informs us,

“all models are simplifications of the real world” (2017, 2741) – meaning all models involve some assumptions and disregard or ignore some details to make the problem of reality manageable. The assumptions made in this model are that we know the composition of the specific MSAMS’s TERs 48, we assume that it operates in 48 Army Recognition Group, “Buk-M3 Viking 9K317M SA-27 Gollum”, https://armyrecognition.com/military-products/army/air- defense-systems/air-defense-vehicles/buk-m3-9k317m-medium-range-air-defense-missile-system-technical-data-sheetspecifications-11312154, accessed 12 September 2022. EXPLORING DECISION ADVANT AGES | 149 a comparable way and uses similar tactics as any equivalent MSAMS would. In a similar fashion, the simplifications made in this model are that it does not consider theatre ballistic missiles 49 (TBMs), nor does it consider other types of SAMS. This research has drawn upon subject matter experts to understand the real problem Subject

matter experts (SMEs) on surface to air missiles and tactics at the Air Defence Regiment (LV6) and LvSS in Halmstad were engaged 50 early on in the conceptual phase to understand the real problem better, and to ensure that the model could either include these considerations or to make informed simplifications of them. One of the most vital questions asked was their most important considerations and variables when deciding the geographical location of a MSAMS. Their consolidated knowledge made it possible to distinguish three key variables: (i) Line of Sight (LoS) towards incoming threat from the air; (ii) the ability to protect whatever the MSAMS is tasked to protect; and (iii) the ability to communicate within the MSAMS unit. The SMEs also made it clear that a typical MSAMS operates in the same way and uses largely the same tactics irrespective of the military organization or nationality. This general inference led to the assumption that a MSAMS operates within a 2,5 km radius: and

the various units within a MSAMS are found within this radius. This is due to the inter-communication link within the system Furthermore, MSAMS are largely used to protect important assets vital to its own forces, such as bridges, airports, or seaports. Lastly, the target engagement radar TER (sometimes referred to as target acquisition radar (TAR) but used as TER in this project) requires LoS to the target to be able to detect, accurately identify and feed information to the missiles. The SMEs were subsequently asked to support the verification of the results in order to fine-tune the model, which is further discussed in the coming sections. Additionally, the MSAMS TERs are mounted on vehicles that require access to roads for movement and operations. The experiment – a concrete example of the military problem The study employs a model designed to be applied as intelligence support to targeting. It is a novel approach where AI estimates and predicts the location of highvalue targets

to augment the decision-making of intelligence collection and optimize the use of resources to find these targets. There are three critical elements to this experiment: first, to construct the conceptual model in an optimal way, which means ensuring that the vital functions 49 As TBMs attack pattern is more or less vertical (overhead), the topography will not restrict any TER’s LoS. 50 A full-days meeting was held at LV6 on the 16th of November 2022, followed up by correspondence via e-mail. The objective of the meeting was to introduce the experiment and receive relevant input to the real problem. 150 | EXPLORING DECISION ADVANTAGES needed to solve the problem are identified, and can be resolved; second, that a complete precursor-model which uses a statistical program 51 to compute signal attenuation of RF-signals can produce synthetic data to create a proper training environment for the neural network to be trained and tested against; and third, that the neural network can

be trained and validated. The performance index for evaluating the model is determined by four criteria: (i) the ability of the neural network to actually mimic and learn from the precursormodel’s statistical program used in the training phase, (ii) the ability of the neural network to improve its performance along the axis of iterations, (iii) the accuracy of the model’s output, and lastly (iv) the ability to reduce the time in completing its task compared to a statistical program. Moreover, the evaluation of Model 1’s performance also uses a reference group of human subject matter experts at the Air Defence Combat School (LvSS) as an external source. The overall objective of this experiment is to develop a computer model that can augment human decision-making in sensor allocation, or even in dynamic targeting assignments 52. The scope of the precursor model is to fully complete all parts of the problem, while the neural network is aimed at showing how a neural network can

outperform a statistical program in that specific function of the complete solution. The results aim to indicate how the use of deep learning can enhance intelligence support to targeting with predictive analysis and decision support. The experiment is also thought to increase the understanding of the implications of using neural networks to make decisions. To the project’s best knowledge, this is the first study in war studies investigating a solution that uses neural networks and semantic segmentation to interpret the topography in satellite images to predict the location of high-value targets (MSAMS), based on the targets characteristics and tactics. The neural network used in this experiment is an exploration of a novel method that enables faster decisions without loss of accuracy. It attributes a semantic meaning to geospatial regions based on the RF signal propagation loss of a MSAM fire engagement (or target acquisition) radar given a specific terrain topography, and the

defined specifications of the radar. Semantic segmentation is a technique applied in computer vision to classify and segment images into different regions based on their semantic meaning. It involves assigning a label to each pixel in an image based on its content, which can then be used to understand and analyze the structure and composition of the scene being observed. Semantic segmentation can be applied to terrain and elevation data in order to improve our understanding of the terrain topography. This type of data is challenging to analyze due to the complex and 51 The statistical program referred to and used in the experiment is Splat! which is commonly used to mathematically calculate signal attenuation of RF based on topography, given a set of RF-parameter. 52 Theoretically, the model could be used to initiate a Strike assignment, for instance a task given to a fighter jet to go directly to the location proposed by the model to be the most likely position of a SAMS target

acquisition radar, and that this would reduce the total time of target engagement given the level of confidence and trust that a human decision-maker would have in such decision support system (DSS). This could be part of a targeting decision policy EXPLORING DECISION ADVANT AGES | 151 diverse range of terrain features that can exist, including mountains, valleys, water bodies, forests, and more. The use of pretrained neural networks can significantly speed up the training process and reduce the amount of training data needed to reach a desired performance of the neural network. Both U-Net and ResNet are pretrained networks that has been used In this experiment. They are both powerful architectures for semantic segmentation and classification respectively in terrain and have been used in numerous studies to extract useful information about the terrain topography (Abdollahi, Pradhan, and Alamri 2022; Fan et al. 2023) This experiment has a spatially aware U-Net architecture designed

for segmentation tasks utilizing a ResNet34’s extraction capabilities for classification tasks as a backbone for emitter location determination. Although semantic segmentation of satellite images has been applied to different problems, the specific method used here is novel. Computational implementation and solution of the problem This section covers the experimental setup including how the model is built, trained, and evaluated. It outlines the conceptual modelling of the problem, the data used and the training and test environment, largely following Robinson’s four basic phases (2017, 2740). The idea of using data-driven insights from a neural network capable of learning signal attenuation of RF-signals, comes from a logical reasoning about RF signals in general. A radar needs to ‘see’ This means MSAMS TER’s must be positioned at a location that enables LoS to a potential incoming threat from the air. The construction of the model must prioritize two key perspectives

throughout the model building phase: the MSAMS and the aggressor or threat. Building a model that can predict the location of a MSAMS necessitates interchangeably shifting the two perspectives. At the center of the model is the definition of the effects that the surrounding topography has for signal propagation losses, given a TER’s specific parameters, its site location, and the threat composition. The topographical effects can be calculated manually using specific tools and traditional formulas to calculate at what distance an aerial threat could be detected by a MSAMS unit from any given place 53. Another common way is to have precise computations made by computer programs, for instance Splat! that compute the topographical effect on RF signals based on specific RF models, such as the Longley-Rice model (see p.121) However, this research proposes using neural networks to make inferences of the topographical effects to RF signal propagation loss. However, the neural network

requires an environment in which it could learn to make such inferences. It necessitated the use of software like Splat! and Splat!HD (high definition) to compute examples for the supervised learning of the neural network. Moreover, the complete 53 FM ‘Handbok Luftvärn Grunder (2020) Arbetsutgåva’, FM2020-22953:2, p.132–133 152 | EXPLORING DECISION ADVANTAGES system had to comprise more functions. These are briefly described in the following section and a detailed system description (Fig. 17) is provided with all its internal and external relations, and dependencies. The role of the neural network is to replace the time-consuming computational part (Step 5 below), also seen in the end of the system description (Fig. 17) and termed ‘compute coverage’ The term refers to the part of the process where most of the computational time is spent and where the majority of the inference happens. The basic architecture comprises of an input of parameters, tasks performed by the

model and an output. The input parameters include a confined geographical area of interest, the parameters of a specific MSAMS TER, and the expected threat height of incoming air attacks. The model then performs a series of subtasks, numbered 1 through 6 below, to solve the problem and generate an output. The solution is a heatmap 54, including specified likely locations of TERs. The main subtasks of the model are described and exemplified below to support a basic understanding: 1. Find all ‘protected’ objects and select them as nodes. Example: Find all bridges 55 within a geo-rectangle. (Fig12) Figure 12: In Step 1 the model finds all objects relating to the query and selects these as nodes. In this example only one bridge was large enough to pass the selection criterion. The selected bridge is found in the center 2. Create Incoming Air Attack Vectors (Fig. 13) 54 In this project a heatmap is defined as a map that depicts values for the main variable of interest (signal

attenuation) indicating the value of the main variable in the corresponding cell range. A legend is placed at the left upper corner of each heatmap defining the corresponding signal losses in dB. 55 Information given by SMEs on the 16th of November 2022 – large bridges are examples of what could constitute a viable protected object for a MSAMS during an operation. EXPLORING DECISION ADVANT AGES | 153 a. b. Set a radius of 2500 m from each node (Reference: SMEs’ third criterion) Create a circle with 5000 m diameter for each node (bridge) and create 360 potential transmitter locations (Incoming Air Attacks 56 ) at 100m height tilting downwards to the bridge center point and spread evenly (one degree separation) around the circle circumference. For each of the 360 transmitter locations create a line-of-sight path between the bridge center point (plus antenna height of radar), and each transmitter. Figure 13: The model creates evenly distributed possible vectors. In this

example we used 64 vectors (5,5degrees in-between), later to be changed to 360 vectors for increased accuracy. 3. Filter Incoming Air Attack Vectors (Fig.14) a. Remove each line-of-sight path between the bridge center point and any transmitter that is obstructed by terrain elevation (topography) 56 The model does not ignore which direction the threat comes from. Instead, it assumes that the threats could come from any direction (360 degree approach). If, however, one would know the direction of the threat beforehand, it would be possible to change these parameters in the coding. This would greatly reduce the time of computation Nonetheless, that would be on a case-bycase basis Additionally, since the TER represents the adversary, any assumption to this come with a risk 154 | EXPLORING DECISION ADVANT AGES Figure 14: The example illustrates how most of the vectors have been filtered out. 4. Find TER target candidates (Fig 15) a. b. For each remaining line-of-sight path, find

highway intersections, nominate each intersection as a potential TER target candidate. Mark each TER target candidate with a unique number Figure 15: The picture illustrates how the model finds intersections of LoS and roads, that it nominates as potential TER target candidates, each given a unique number. EXPLORING DECISION ADVANT AGES | 155 5. Compute the signal attenuation at each nominated TER target candidate a. b. 6. Recalculate 2b for each nominated TER target candidate. Normalize the results of all calculations as a transparent RF power Heatmap visualizing the relative opacity of all individual layer resulting in a composite heatmap with each TC displayed. Generate a Heatmap revealing the results (Fig. 16) Figure 16: The picture illustrates the output of the model. Target candidates are clearly marked. The signal attenuation shows 0 to -150dB attenuation for a simulated SA-27 Target engagement radar. Note: The neural network is a replacement for the computations made

in Step 5. The training and test environment was a synthetically created supervised learning environment from building the baseline model that uses computational software in Step 5. Establishing the baseline model was essential to enable the supervised learning of the neural network to perform semantic segmentation on satellite images and emulate how topography affects signal attenuation of RF signals. 156 | EXPLORING DECISION ADVANTAGES Figure 17: A detailed system description of the model, with the loops and sequencing of subtasks (Step 1-6). EXPLORING DECISION ADVANT AGES | 157 The system description (Fig.17) is the architecture behind Model 1 The complete script of instructions (algorithms) used is perceived to be of lesser value, as the processes, loops and the sequencing of subtasks are presented in Fig.17, thereby portraying the construction of the model, aiding transparency and explainability to the project. At the top of the system description four software packages

(Main program, overpy, polycircles and Wavetrace) are correlated to their respective engagement in the subprocesses. The two main loops are seen in the two rectangle shapes, ending with the task of normalization and the subsequent presentation in Google Earth Pro. Training the neural network with a synthetic ‘tutor’ – a novel method of radar emitter location determination A training environment was created for the neural network. Prior to training the neural network to predict RF signal attenuation based on topography, a synthetic ‘tutor’ had to be created. This tutor was built based on a combination of parameterization of synthetically created radar emitters, terrain images, and a computer program to generate radio signal coverage reports to an annotated topographic map depicting the expected coverage area of the radar emitters. This led to a supervised learning consisting of terrain images and corresponding labels for these images. The basic architecture consisted of an

Encoder (down sampling path with ResNet34 backbone), a Bottleneck of additional convolutional network layer (to process small spatial features), a Decoder (up sampling path) and an Output layer to map the highresolution features. What follows is a more complete disclosure of how the environment was set up, to include the training process, test and evaluation process (chronological steps from A-D). A. Generating the foundation – grids and antenna locations The workflow for the training process starts with computation of the radio signal coverage reports for the transmitters of a MSAM TER. A combination of a custom written grid generator is used for generation of a set of evenly spaced-out radar locations across a latitude/longitude grid, combined with Wavetrace 57 software for generation parametrized radar antennas positioned along the defined grid. For calculating a regularly spaced grid in meters, the research projects the latitude/longitude boundaries of the selected area into a

coordinate system that supports distances in meters. This is done by transforming the area of interest into the spatial reference standard EPSG:3035 58. It then creates a grid of evenly spaced points by iteration across that area and then projects the points back to 57 Wavetrace is a Python 3.5 package designed to produce radio signal coverage reports 58 Spatial Reference, “EPSG:3035”, https://spatialreference.org/ref/epsg/3035/ , accessed 12 September 2022 158 | EXPLORING DECISION ADVANTAGES latitude/longitude pairs as visualized in Figure 17. These synthetically created radar locations are then used as a basis to parametrize a MSAMS TER on each location and create a CSV file 59 with the all the parameters. For emitter specific parameters TER data is used. The chosen TER is the type used in the Buk-3M medium-range mobile air defense missile system 9K317M 60 , also referred to as SA-27 GOLLUM 61 (see Figure 18). Figure 18: The Buk-M3 MSAMS 9K317M, referred to by NATO as

SA-27 GOLLUM, Source:armyrecognition.com The parameters have been collected from open sources and adapted to Splat! 62 format. It considers antenna height, radar frequency, power output, polarization, bearing, horizontal and vertical beamwidth and the level of down tilt of the radar. The result is a csv file containing 1,911 evenly spaced-out radar transmitters in a grid pattern defined by two diagonal latitude/longitude boundary points (Fig. 19) 59 A comma-separated value (CSV) consisting of data separated by commas. 60 Missilry Info, “9K317M "Buk-M3" medium-range anti-aircraft missile system”, https://en.missileryinfo/missile/bukm3, accessed 12 September 2022. 61 Army Recognition Group, “Buk-M3 Viking 9K317M SA-27 Gollum”, https://armyrecognition.com/military-products/army/airdefense-systems/air-defense-vehicles/buk-m3-9k317m-medium-range-air-defense-missile-system-technical-data-sheetspecifications-11312154, accessed 12 September 2022 62 Splat! is an RF Signal

propagation loss, and terrain analysis tool for the electromagnetic spectrum between 20 MHz and 20 GHz, based on U.S Geological Survey and Space Shuttle Radar Topography Mission (SRTM) elevation data The tool is used extensively in research. See: Splat!, “Introduction”, https://wwwqslnet/kd2bd/splathtml, accessed 5 November 2023 EXPLORING DECISION ADVANT AGES | 159 Figure 19: Visualizing the method of creating evenly spaced-out antenna grid generation of radar emitters. B. Generating training data – preprocessing images and labels To generate the training data, the research uses a set of input terrain images along with the corresponding labels for these images. The labels are created by using the radar parametrization file (described in A) as input to Splat! for generating radio signal coverage reports to each entry in the CSV file. For each CSV entry, Splat! outputs an annotated topographic map depicting the expected coverage area of each radar transmitter. The coverage

area is defined in decibel attenuation values in 16 levels, from 0dBm to -150dBm in 10dBm decremental steps, as depicted in Figure 20 below. The project uses the PySplat Python wrapper to generate the training data 160 | EXPLORING DECISION ADVANT AGES Figure 20: The decibel attenuation values in 16 levels, from 0dBm to -150dBm in 10dBm decremental steps. The pre-processing stage consists of rectification and cropping. Each training example consists of a before and after part. The before part is represented by the raw SRTM 1 tile (satellite images) without the radar signal coverage calculation. These satellite images were sampled from the SRTM-1 63 dataset covering most of the USA central land mass terrain. A choice was made to use this terrain since it included a variety of terrain types, which was beneficial from a training perspective. The before part therefore consisted of 1,911 evenly distributed antenna locations in a grid with 50km squares. The after part consists of a radio

signal overlay generated by the neural network. See Figure 21 and 22 for a training example consisting of a before and an after image. The training examples containing the before and after pictures had to be rectified in order to match each other in a pixelwise comparison. A custom written Python code was used to align the pictures pixelwise and crop 64 them accordingly (remove unwarranted areas), to obtain similar resolution of 1024 by 1024 for each picture. 63 The Shuttle Radar Topography Mission (SRTM) was completed by the space shuttle Endevour in 2000. SRTM collected radar data over 80% of the Earth's land surface between 60° north and 56° south latitude. The SRTM 1 dataset offer worldwide coverage of void filled data at a resolution of one arc-second (30 meters) and provide open distribution of this global data set. Source: NASA, “Shuttle Radar Topography Mission 1-arc second Global”, https://cmr.earthdatanasagov/search/concepts/C1220567890USGS LTAhtml, accessed 5

November 2023 64 This process ensures that all unwanted areas of the images discarded. EXPLORING DECISION ADVANT AGES | 161 Figure 21: The before part; a raw terrain image without the radar signal coverage calculation 162 | EXPLORING DECISION ADVANTAGES Figure 22: The after part; the terrain image with a radio signal coverage overlay C. Training the neural network A U-Net was used with a Resnet34 65 as the backbone for the training. U-Net is a stateof-the-art image classification model, pre-trained on the ImageNet dataset 66 which contains 100,000+ images across 200 different classes. The ImageNet is a collective research effort by Stanford University and Princeton University to provide researchers with image data for training large-scale object recognition models. The research finetuned the neural network on 1,380 of the training batches generated in three sessions of approximately 400 examples in each batch (75% used in training and 25% used for validation) as visualized

in Figure 23 from top to bottom. 65 The architecture is implemented from the paper ‘Deep Residual Learning for Image Recognition’ by He et al. (2015) 66 Stanford Vision Lab, “ImageNet”, https://www.image-netorg/, accessed 5 November 2023 EXPLORING DECISION ADVANT AGES | 163 164 | EXPLORING DECISION ADVANT AGES Figure 23 Comparisons between the four batches. The top picture through to the bottom picture shows how the prediction by the neural network is improving along the axis of iterations – it learns and becomes increasingly better in its predictions. While the first batch showed resemblance in the predictions, the second and third training run revealed visible progressions to the neural network’s learning curve. The fourth batch confirmed that the network was behaving as intended, reaching a level of prediction that was thought to be enough. To evaluate the performance of the neural network and the model as such, model as such, this was done using commonly used

performance criteria and simulations Results Evaluating of performance of the model When evaluating the performance of deep learning in object detection and image segmentation tasks, a commonly used metric is intersection-over-union (IoU) (Berman, Rannen, and Matthew 2018; Rahman and Wang 2016). The model accuracy is measured after each training batch (Fig.23) The IoU does a pixel-bypixel analysis and calculates the ratio of the overlap to their respective combined ���� areas; we can write IoU as (������) = whereas TP= True Positive; FN=False ����+����+���� Negative; and FP= False Positive. The IoU scoring ranges from 0 to 1 While 1 is perfect in theory, more than 0.7 can be considered good as seen in the straightforward object detection example (Fig. 24), that is referenced in this context to better understand this correlation metric. However, the example given should be perceived in the light of each pixel, meaning that the neural network

has to learn to do its image segmentation task from the satellite image at a pixel level, so that it has an accuracy at a pixel level that is similar to the middle image in Fig 24. Given that each image had a resolution of 1024 by 1024 each image will have over one million pixels in total. EXPLORING DECISION ADVANT AGES | 165 Figure 24: Example of IoU metric. The left has the highest IoU scoring since the predicted box almost coincides with the ground truth box. Hence, as seen in the other examples, the less they coincide, the lower the scoring. Source:https://www.v7labscom/blog/intersection-over-union-guide Regardless of the origin of the ground truth data (in our case NASA’s SRTM-1 data), or how careful the labelling is done, it is unlikely to have the predicted output completely matching the ground-truth bounding box coordinates. The accuracy of the neural network from the four batches was verified using IoU metric (Table 1): TABLE 1 Measurement of intersection-over-unit,

IoU Batch IoU scoring 2 0.72 1 3 4 0.53 0.76 0.77 As described in Table 1, the IoU scoring increased along the axis of training. It also became clear that, in order to improve the IoU scoring beyond the results after the fourth batch, it would likely have required an increasingly larger volume of training examples, which was considered unwarranted for the purpose of the study. Apart from adding more training data, another option could have been to try to configure the training data by changing the angel of viewing. However, this was never done Another metric used for measurement was time. The research propagates for speed as this can be an asset in the quest for effective targeting and fast response time 166 | EXPLORING DECISION ADVANT AGES when providing intelligence updates and the ability to swiftly act in a dynamic environment. It therefore compares the inference time for the neural network inference in contrast to calculations made by Splat! using the Longley–Rice

radio propagation model to predict the attenuation of the radar signals. Three batches of increasing sizes of terrain examples are measured for each method to assess the speed performance. Table 2 and 3 depict the time it takes to generate radar coverage reports for three batches of terrain examples for each method. The Batch size is over 10, 100, and 1000 examples, respectively. TABLE 2 Splat! Calculation time Batch size Time(s) 100 2694 10 1000 203 25993 TABLE 3 ResNet34 Inference time Batch size Time(s) 100 65 10 1000 7 639 At face value, the evaluation of the neural network reveals that the ResNet34 inference time is roughly 40 times faster than Splat! calculation time. Given that approximation, what takes Splat! almost an hour to calculate, will be finished by ResNet34 in just over a minute. If time is of the essence, any model used in support of dynamic targeting would have to be fast in its prediction. The result is promising as the model has magnitudes of

performance increase when considering inference versus calculation time and the model is able to generate incrementally better EXPLORING DECISION ADVANT AGES | 167 results along with an increased number of training examples, but the research suggests that more training examples are required to exploit the full potential of this new method. Nevertheless, even if time can be reduced by 40 times when compared to one of the more frequently used calculation models (Splat!), the task completion of ResNet34 necessitates more computer memory than available. Even if the laptops used were advanced 67, more GPU 68 is required, involving other solutions equipped with GPUs equivalent to NVIDIA A100 69 . In scenarios where time is not of essence, such as performing predictive analyses prior to an operation, or in support of concept development, modelling and simulations or exercises, then neural network would likely still be preferred due to its ability to transfer its learning to any other

satellite imagery that has the same image resolution. Simulations - Applying the model beyond its training data The model was applied in different simulations to validate its performance. Though the simulations shared the same objective of predicting the most likely location of a SA-27 GOLLUM Target engagement radar (TER), they differed in their respective geographical areas. The areas were selected based on their variation in topography From a methods perspective, the differences between the topography in each simulation were the most important criteria as the model’s ability to manage variations in topography is considered the most critical ability. However, the selection was constrained by SRTM-1 land coverage (between 60 degrees north to 59 degrees south of the equator). This made it hard to find more mountainous terrain in Sweden. A second criterium was also used in the selection of the geographical locations. As SAMS, from a tactical perspective, are used for protecting vital

assets for a joint force, to include bridges and airfields, it seemed reasonable to use these types of infrastructure as points of reference. In the end, the selected areas became (i) Simlångsdalen Bridge at 56°43'10"N 13°08'38"E in Sweden; (ii) Skiathos Airport at 39°10'36"N 23°30'12"E in Greece; and (iii) Kveitseid Bridge at 59°21'48"N 8°31'21"E in Norway. Target candidates in all simulations share the same vital attributes to the model. They all have high relative radar coverage to enable a TER to detect any aircraft, drone, or missile coming in vicinity of their vectors at 100m height. Moreover, they are all accessible by roads that can accommodate a TER vehicle. 67 The laptops used were Dell Precision Corei9-11950H with 8 Core, 128GB RAM, and NVIDIA RTX A5000 w/16 GB GDDR6. 68 Graphics processing unit (GPU) is a specialized processor in a computer designed to accelerate graphics by processing parallel work

simultaneously . Its ability to do so also makes GPUs useful for applications like deep learning 69 NVIDIA is a company producing state of the art GPUs for deep learning such as the A100. For measurements in this, referenced benchmarking, see for instance tests as of September 2024: Bizon, https://bizon-tech.com/gpu-benchmarks/NVIDIA-RTX-3090-vsNVIDIA-A100-40-GB-(PCIe)/579vs592, accessed 2 September 2024 168 | EXPLORING DECISION ADVANT AGES (i) Results from the first simulation at Simlångsdalen Bridge (Fig 25). The topography was fairly flat in this area. The terrain was thought to be good from a LoS perspective which was verified and elucidated by the colors representing dB (see scale in Fig. 20) Eight target candidates (TCs) were suggested by the model. The limited number of roads narrowed down the number of target candidates. Figure 25: The output from the first simulation - Simlångsdalen Bridge. E (ii) Results from the second simulation at Skiathos Airport (Fig. 26)

This area was chosen because of its delimited options in plain sight. The higher mountainous areas in the north-northwest were beyond the 2,500 meters radius EXPLORING DECISION ADVANT AGES | 169 (some ridges being almost 4,000 m from the protected object). To illustrate this, Figure 26 shows both a traditional Google Earth Pro picture on the top and the model’s outcome at the bottom. This enables the viewer to compare the two The cut-off areas are cropped because no target candidates were nominated (only sea within the range of 2,500meters). 50 target candidates are suggested by the model, with a few in the north at the border of the range-limit, suggesting that if the range had been extended these target candidates would have been even further out. Figure 26: The combined image shows both a traditional Google Earth Pro picture on the top and the model’s outcome at the bottom. This enables the viewer to compare the two. An additional picture is provided to illustrate the

importance of accessible roads. The model only considers target candidates that have the ability to reach an area by vehicle (Fig. 27) 170 | EXPLORING DECISION ADVANTAGES Figure 27: Image viewed towards the southeast showing how the model has nominated some target candidates whilst the higher ridge to the right in the image has no candidates due to the lack of roads. (iii) Results from the third simulation at Kveitseid Bridge (Fig. 28) The topography was varied in more than three quarters of the total area. The number of target candidates suggested by the model was 43. Several minor roads made the more mountainous areas accessible. Figure 28: The output from the third simulation. The heatmap is displayed on Google Earth in this picture and has been tilted to display the result in three dimensions. EXPLORING DECISION ADVANT AGES | 171 External validation (web survey) and fine-tuning of the model Evaluation is a critical part of experiments as they convey a bridge to both

validity and reliability. The evaluation of the model (and the experiment as such) was done by revisiting the four criteria of the performance index and the overall objective of the experiment, starting with the former. The first performance criterion was the ability of a neural network to mimic and learn to emulate signal attenuation of a TER radar. By creating a synthetically generated training environment using satellite images from SRTM-1 and some commercial programs for statistical calculations of RF signal propagation loss due to topography, the neural network was able to be trained and validated (Fig. 23 and Tables1-3). The second and third criteria were also met, since the U-Net improved its ability to emulate and infer along the axis of iterations (Fig. 23), and the accuracy of the model’s output could be verified (Table 1). Lastly, the fourth criterion of the performance index was the ability to reduce the time compared to a statistical program, which was verified (Table 2

and 3). In addition to these four criteria, the evaluation of Model 1’s performance was also to be validated by using an external group of human subject matter experts. This was done by giving five SMEs an individual task of suggesting the preferred site location of the same type of TER, in the same areas that were used for simulations, and subject to the same conditions that the model was given. They were initially informally asked if they could support this evaluation stage, to which they agreed. The tailored task was done individually and anonymously through an online web survey during December 2023 (see Appendix A) The task was conditioned as: Choose the locations of two TERs in the three given geographical areas. The TERs has an antenna height of 5 meters, and the type is the equivalent to the Buk-3M. The unit is a medium-range, vehicle-born SAMS consisting of 2 Target engagement radars (TERs), 2 Surveillance radars, and 6 Missile Launchers. The mission’s essential task for

the unit is to protect either an airport or a bridge. The threat is 360 degrees and consists of fighter jets or missiles released from fighter jets, with an altitude between 30-100 meters. To solve the task, you can divide each area into four quadrants of 90 degrees each and select 2 sites per quadrant. The suggested locations are subject to two constraints; maximum distance to the protected object is 2,500m and the location must be accessible by vehicle +- 50 meters. (Translated from the Swedish text given to the SMEs, Appendix A) The SMEs used Google Earth to work through their respective solutions and marked their suggested site locations according to the instructions. If they found an optimal location, they were to name this ‘optimal,’ and to give it common attributes (the 172 | EXPLORING DECISION ADVANTAGES reasoning behind) related to tactical concepts and surface-to-air instructions 70 . When consolidating the answers (Fig. 29), the most frequently used attribute was

that the optimal location was perceived as ‘appropriate from a targeting perspective’ (‘Lämplig ur ett verkansperspektiv’ 71 ). The language used is Swedish, but most tactical attributes include an English translation in brackets. The last option, 'Other' ('Annat' 72 ), is optional and serves as a 'Supplementary Information' field for explaining specific details related to the motivation behind the choice. 70 FM ‘Handbok Luftvärn Grunder (2020) Arbetsutgåva’, FM2020-22953:2. 71 Swedish phrase as used in the survey. 72 See note above. EXPLORING DECISION ADVANT AGES | 173 Figure 29: Distribution of attributes from the SMEs three Optimal TER site locations. 174 | EXPLORING DECISION ADVANTAGES Most importantly, the use of an external group of human SMEs made it possible to superimpose their respective suggested solutions onto our model’s solutions, and to compare the results for evaluation purposes. Figures 30 and 31 illustrate two

of the simulations where the SMEs suggestions are superimposed into Model 1’s predictions. Since there were five SME respondents, three geographical areas, and each respondent suggesting one-two optimal site locations per area for each respondent, the total amount of suggested locations were in theory 15-30. Each location is marked using Google Earth, which means that the suggested or predicted location in Google Earth has a point precision of a few meters, according to recent research (X. W Wang and Wang 2020, 1054) Each geographical area in the simulation has a radius of 2,500 meters from the protected object, rendering a total area of approximately 20km2. Arguably, being within about 300 meters of any of the SMEs’ defined locations is inferred as being at more or less the same TER site location, as that is the equivalent of just over 1 percent of the total area. Figure 30: The image is a portion of the Kveitseid Bridge simulation area. The blue marker represents a respondent

suggestion, and the yellow marker is a prediction by the model. A ruler is inserted to measure the distance between the two The distance is 60 meters. Note the white road (gravel road) in the image making the sites accessible by vehicles. EXPLORING DECISION ADVANT AGES | 175 Figure 31: The image displays another vector of the Kveitseid Bridge simulation area. The blue marker represents a respondent suggestion, and the yellow marker is a prediction by the model. A ruler is inserted to measure the distance between the two The distance is 300 meters. Note the white road (paved road) in the image making the sites accessible by vehicles. Also note that the suggested site is likely to have been closer if the respondent had had access to this heatmap. After reviewing the model's predictions with the SMEs, the general impression is that there are more similarities than differences. The most significant difference was in the way that the SME’s chose site locations close (within

300m) to the protected object. Since the model used highly sophisticated data to calculate the best possible site location along any of the 360 vectors that still had LoS after the filtering (Step 3), it ‘knew’ to mark the target candidate at a more feasible location than the SMEs as seen in Fig. 30 Though approximative, having only Google Earth as a tool, the SME’s individual expertise and suggested solutions were helpful to support further fine-tuning. More importantly, the similarities (Fig 30 and 31) are indicative of good predictions by Model 1. The post-survey analysis also enabled a subsequent finetuning of the model, in which the model was given an additional task to optimize the nominated target candidates it provides in Step 3. In this add-on the model considers the balance (K-means clustering) of the cluster size relative to the minimum distance between each cluster and each reference point (being the human experts’ input). 176 | EXPLORING DECISION ADVANT AGES

Additional fine-tuning of the normalization and visualization on Google Earth Pro Additional fine-tuning of the normalization and visualization on Google Earth Pro was done at a later stage to try and increase the transparency of the results. As this was performed after the actual simulations were conducted, only one samples is added here (Fig. 32) The figure is remarkably more transparent, revealing more of the Google Earth Pro’s map features. Figure 32: A sample of how the results can be visualized after fine-tuning, enabling more transparency in the image. The sample shown (Fig. 32) is after some fine-tuning clearly revealing more terrain, buildings and roads that is within all figures provided in this experiment, though not as apparent as in this image. Moreover, the three suggested TER locations in the image (Fig. 32), show how the model is using the constraints given in the instructions to only consider locations that are accessible by road, as the TERs are vehicle-borne.

Discussion The last section of the chapter discusses the experiment, the results, and the implications of Model 1, including an analysis using the operationalized theoretical tool (Table 2 in Chapter 3) and a concluding section over the chapter. EXPLORING DECISION ADVANT AGES | 177 Considerations of reliability and validity The most important factors when considering the reliability and validity of the experiment are arguably: (i) the interaction with the SMEs to understand the real problem prior to the model building; (ii) the selection of three different topographical areas for the simulation phase; and (iii) the validation of the predictions (output of the model), made possible through the survey, after which the respondents’ suggestion could be superimposed and compared with the model’s predictions. The three factors strengthen the validity of the experiment, the reliability of the results, and relevance (or utility) of the model. Limitations The experiment did not

explore the effects of vegetation or weather, which impacts on how radar signals propagate, as suggested by Skolnik (2001, 482–83). Nevertheless, the research indicates that at least vegetation would be possible to include, since neural network architectures allows for the integration of additional information at a pixel level when performing semantic segmentations of images. Implications The overall objective of this experiment was to develop an AI model that could augment human decision-making in sensor allocation, or even in dynamic targeting assignments. The application (model) recommends Target Candidate Sites for Medium-Range SAMS based on rationality (human expert input), and logical conditions (parameters and signal propagation loss due to topography). The model can support decision-making during target development where it designates target area of interest (TAIs) of adversary’s high value targets (HVTs being the Target Engagement Radar Sites (TERs)). The results

indicate how it can be beneficial for a joint force to use similar tools in more situations other than dynamic targeting. However, before addressing other more subordinate benefits from this experiment, the discussion first explores its intended role as augmenting decision-making for efficient sensor allocations, and its potential role to initiate a strike assignment during dynamic targeting. To do so, the operationalization found in Chapter 3 is utilized as the baseline for this principal analysis. 178 | EXPLORING DECISION ADVANTAGES Analysis Table 3 The table provides an overview of the different tasks (activities) that Model 1 contributes to, as well as its relational engagement within the JTC and the OODA loop. OODA Observe JTC Phase 1 Dyn TGT Find Cmr’s Intent, Fix Phase 2 Orient Phase 3 Capability analysis Track Phase 5 Mission planning # 1 Search Indicators & warnings 1 3 Identify 2 4 6 Target Detect JIPOE Determine TGT characterization

Target System Analysis Classify Confirm TGT validation Significance 9 TGT mensuration 8 11 3 Target Material Production 5 4 6 PID Estimations 7 Prioritization Intermediate target development 8 Monitor ISR management 12 Estimate TGT window of vulnerability 14 Desired effects Functional Risk estimation 15 Time management 16 RoE/TGT Policies 17 Collateral damage Estimation (CDE) 19 Options 18 2 MoP/ MoE Basic target development 7 13 Cmr’s decision, Decide Intel support activities 10 Phase 4 Force assignments Subtasks/Functions 5 Objectives, and guidance Target development # Threat analysis characterization 9 10 Damage estimation 11 Functional analysis including Advanced target development 12 Threat analysis 13 EXPLORING DECISION ADVANT AGES | 179 20 Recommendations 22 Deconfliction/Coord/ Synchronization 21 23 24 25 Force execution Engage Exploit Act Assessment Assess 26 27 28 Requirements incl Combat

assessment Intel requirements 14 Collection tasking 16 Combat Assessment Battle damage assessment (BDA 1) 17 Re-attack recommendations Re-attack recommendations 19 Select option CID Field CDE Target engagement Measurements & Documentation Follow-on actions 30 Re-attack 32 MEA 33 15 Issue order 29 31 Sensor allocation BDA 2 BDA 3 18 20 When reviewing the collection of intelligence support activities that have been derived from the targeting doctrines (Fig. 5), the subsequent interpretations of the dynamic targeting (Fig. 8), and the OODA loop, it is suggested that the yellowmarked areas in Table 2 corresponds to the outcome of the experiment put forward in the chapter. Beginning at the top with activity #4, Target System Analysis (TSA), this has been established as a vital activity and product as Chapter 2 and 3 have revealed. The model developed (Model 1), supports the efforts of establishing a foundation for target development from building knowledge

of an adversary. More specifically, the model can be applied to questions within a TSA or used to provide tailored target material productions (#5) or any intelligence estimation (#7). Large geographical areas will be covered using neural networks to support intelligence production and 180 | EXPLORING DECISION ADVANT AGES decision-makers with data-driven insights, as indicated in this study. Additionally, knowledge produced by the AI model can be cross-referenced with previous knowledge about the adversary. This can be arranged as a feedback loop – a constant update of the TSA or other products before, during, or after an operation. The research propagates for speed and believe the model can be a valuable asset in the quest for effective targeting and fast response time when providing intelligence updates. Although the ResNet34 outperforms Splat! when comparing time for inference and calculation respectively, this model needs more training examples to exploit the full potential

of a method using neural networks. The model accounts for topography, including any object such as buildings that are represented in an image, but not vegetation. This is, however, only a matter of having that type of training data. The full potential would be a model that accounts for all the existing TERs, or even all the existing equipment using RF signals, which could be valuable for any functional analysis (#12) or threat analysis (#13) of a specific system using RF. The main advantage of the model, however, is found in its support to the three intelligence activities: ISR 73 management (#9), Sensor allocation (#15) and Collection tasking (#16). The three activities all focus on using sensors effectively and efficiently. The first is a pre-planning activity aimed at ensuring the availability of collection resources, while the latter activities address the use of specific sensors for specific tasks. All three activities also share a common value in ISR optimization, and in

enabling a faster target engagement. If the model were to be integrated into dynamic targeting, the result from this study indicates that this would benefit several decision-making aspects. The model can augment and present likely TER sites for MSAMS. Other RF dependent systems could also be incorporated, so that when it is deemed necessary to find targets of significance such as MSAMS, to model could become a vital tool, and reduce the total time of target engagement. Given the level of confidence and trust that a human decision-maker would have in this type of decision support system (DSS), the output of the model could become an input, triggering an initiation of a strike assignment. A strike package 74 could be redirected to head to the location recommended by the model and engage the target. The opportunities offered would presumably necessitate changes to existing practices and targeting decision policies, with a trajectory towards facilitating novel targeting services. This is

why ENGAGE is highlighted The model is assessed as impacting five out of seven of the elements of F2T2E2A 75, where the main contribution is its support to sensor allocation. It is also highly influential within the joint targeting cycle. However, the augmentation is more 73 Intelligence, Surveillance and Reconnaissance Management is the guidance and use of the collection resources. It also incorporates the methods employed in the planning, collection, processing, exploitation, and dissemination of intelligence in military operations. 74 In this context a strike package would constitute any effector or composition of effectors that had the ability to locate, perform a positive identification, collateral damage estimation, and munition that could create the desired effects. 75 Find, Fix, Track, Target, Engage, Exploit, and Assess. Also see Chapter 3 and 4 (Operationalization) for more details EXPLORING DECISION ADVANT AGES | 181 towards enhancing decision-making with predictive

analyses. The more adversarial systems it could integrate, the more value it is likely to have when used for predictions. If complemented with feedback loops from real-world data, such as post-strike assessments, this could be used to further fine-tune the model. The experiment also indicates how the use of deep learning can enhance intelligence support to targeting. It increases the understanding of the implications of using neural networks to make decisions. The intended solution was to build a model that was able to present recommendations on where to allocate own sensors in a dynamic setting, and by this acting as a targeting advisor expounding likely TER sites for a decision-maker. Deep learning-based decision tools adds “more apparent competency, more breadth”, and perhaps “a tendency, as a result, to turn over more decision-making to them”, as stated by Turek, the deputy director of the Information Innovation Office at the US Defense Advanced Research Projects Agency.

76 Furthermore, the solution was to present a new way of making use of neural networks to facilitate a reduction in time. As such, the solution has significant potential to scale, allowing for the inclusion of additional adversarial models, which could improve predictive accuracy and advance preemptive insights. Additional benefits from using neural networks infer effects of topography on RF signal attenuation, can be found in augmenting own air forces tactical target attack planning. Returning to one of the results, but this time highlighting one of the obvious TER ‘blind spots’ with a white arrow towards the protected object (or target from an attacker’s perspective), it clearly defines a potential air attack approach (Fig. 33). In this vector the TERs have little if any ability to discover an incoming threat at 100m height. 76 Michel, Arthur Holland, ’Inside the messy ethics of making war with machines’, MIT Technology Review, 16 August 2023,

https://www.technologyreviewcom/2023/08/16/1077386/war-machines/, accessed 2 September 2024 182 | EXPLORING DECISION ADVANTAGES Figure 33: The outcome of the model, in which the least covered vector is highlighted with a white arrow to indicate a potential air attack route for a drone or a missile. In support of this argument, it can be worth noticing that in Fig. 33, twelve TERs are producing this result, compared to two or three TERs that normally would constitute a similar MSAMS unit. Conversely, the application could also be used to enable less predictive behavior of own forces and assets when integrated in trials, exercises, or defense planning. EXPLORING DECISION ADVANT AGES | 183 Conclusion The experiment covered an AI model built to augment decision-making by recommending the location of high value targets (HVTs). Being able to predict or make inferences of this will enhance precision and efficiency in a joint force dynamic sensor allocation and target engagement. The

research propagates for speed and believes the model can be a valuable asset in the quest for effective targeting and fast response time by providing support to intelligence collection. The experiment indicates how the use of deep learning can enhance intelligence support activities for dynamic targeting. The fundamental AI technique applied to the problem was deep learning (DL). Deep learning is complex, takes time, and requires computational power. In this experiment, its neural network structure was to learn, through semantic segmentation of satellite images, how topography affects signal attenuation for target engagement radars (TERs). This has been proved in other research areas to be highly effective for semantic segmentation tasks (Fan et al. 2023; Abdollahi, Pradhan, and Alamri 2022). The more training a modern neural network is given the better its results become. This was confirmed by the evaluation of the neural network’s performance. The method (or approach) to solve a

problem of emulating the impact of topography to radio frequency signal propagation loss (signal attenuation) is novel 77. The model was built using a combination of U-Net and ResNet34. The results indicate a vast reduction of time, and an improved prediction accuracy, along the axis of more training and available data. Specifically, it permits taking directionality and topographical information into account when determining the radio emitter location. The main advantage of the model is found in its support to ISR management (#9), Sensor allocation (#15) and Collection tasking (#16). These are activities engaged in using sensors effectively and efficiently. They refer to ISR optimization and enable faster target engagement. If the model were to be integrated into dynamic targeting, the result from this study indicates that this would benefit decision-making aspects. The model can augment and present likely TER sites for medium-range surface-toair missile systems (MSAMS). It could

become a vital tool, reducing the total time of target engagement. Given the level of confidence and trust that a human decisionmaker would have in this model, the output could initiate a strike assignment A strike package could be redirected to head to the location recommended by the model and engage the target. The opportunities would likely impose changes to existing practices and targeting decision policies, with a trajectory towards novel 77 The term method is used in this context to correlate to how it is referred to in the patent application no. 2300078-9, 20 September 2023. 184 | EXPLORING DECISION ADVANTAGES targeting services. These type of opportunities or extensions of this research are discussed in Chapter 7 Conclusions. The analysis of the model’s impact to the theoretical framework indicates that it has bearing on five out of seven of the elements of F2T2E2A, and that it is influential to most parts of the JTC and the OODA loop. Perhaps most significantly, in

reference to the OODA loop, the model signifies how foreknowledge of an adversary can be further enhanced. The application of neural networks has some advantages: First, neural networks can operate in millisecond regimes, enabling real-time determination of RF emitter location. Second, their performance improve over time As more predictions are made and compared to real-world data or high-accuracy simulations, the neural network parameters can be updated to enhance performance. Hence, along with more training examples, values relating to an optimal RF emitter location can be determined – and by doing so, be automated, and applied as a real-time determinator of an adversary’s specific TERs. Conversely, an automation could also be fielded for planning purposes, either to explore the optimal RF signal coverage in an area of operation, or to find out the minimum RF power output required to allow for a more cautious and stealthier RF footprint. Applying the method of neural network for

any RF communication planning is thought to liberate communication planners from time-consuming computational software. By training neural networks in these matters, large geographical areas can be covered in an abbreviated period of time illuminating the best routes for communication, or other uses where RF Line-of-Sight (LoS) is of importance. This chapter explored how neural networks and deep learning can be applied to targeting. The intelligent agent (the AI application) learned through supervised training to make predictions. In the following chapter, the intelligent agent will use machine reasoning based on pre-defined rules to optimize the best possible decision given its input data. EXPLORING DECISION ADVANT AGES | 185 186 | EXPLORING DECISION ADVANT AGES Chapter 6 – Experiment # 2 Model 2 - Optimizing a Joint Force’s Dynamic Targeting Decisions with machine reasoning "Remember that all models are wrong; the practical question is how wrong do they have to be

to not be useful." Introduction (Box and Draper 1987, 74) 78 This chapter sets out the second experiment of this thesis that focuses on optimization and dynamic targeting. It frames the field of optimization by giving a general description of optimization as maximizing or minimizing an objective function along with a description of the military problem at hand. This refers to deciding which means to use for which targets to achieve optimal effectiveness. The chapter presents the setup of the experiment, the data set used and the subsequent modelling and simulation, evaluation, and results. It concludes by assessing the implications of the experiment and its results. The experiment explores how an AI can be employed for dynamic targeting purposes and the implications it brings. Model 2 (the specific AI-application) is given the role of optimizing decision-making using machine reasoning under dynamic targeting conditions. Optimization entails the quest for the best or most

efficient solution under given constraints. It offers tools and frameworks to model and solve an eclectic range of problems (Hillier and Lieberman 2021). Machine reasoning has become an increasingly popular research topic in recent years due to its potential to automate tasks and improve military decision-making (Luotsinen et al. 2019) In combination, these two attributes of Model 2 create a transparent technique that is applied to augment decision-making in targeting. Previous research has shown that optimization algorithms can be used to improve the precision when augmenting decision-making related to weapons to target assignment (WTA) problems (Babul 78 The more famous version of this quote is perhaps ’all models are wrong, but some are useful.’ While no model can perfectly represent reality, certain models can still provide valuable insights and guide decision-making effectively, which is essential in optimization. EXPLORING DECISION ADVANT AGES | 187 Hasan and Barua

2021). However, little research has been done to combine multiobjective criteria using a lexicographical order with re-attack options based on the outcome of a previous stage. Certainly, real-life military problems can have many objectives that need to be satisfied, just as in other fields of research. These problems are often classified as ‘lexicographic multi-objective problems’, “where the first objective is incomparably more important than the second one which, in its turn, is incomparably more important than the third one, etc.” (Cococcioni, Pappalardo, and Sergeyev 2018, 298). A lexicographical order (or approach) therefore involves arranging the objective functions in order of importance and solving a sequence of single-objective problems. Each objective or criteria is optimized in order of priority, with higher-ranked objectives (or criteria) taking precedence. This combined approach to the problem also uses realistic data inputs of weapons and targets verified by

subject matter experts (SMEs) from within the SwAF. The results indicate that the model can solve the problem and find an optimal solution, thereby, in theory, supporting the effectiveness of an operation. The results also suggest that it could have utility for cost-benefit analyses, capability analyses or analyses of decision policies. To this project’s best knowledge, this is the first study in the literature to address a solution that qualifies as relevant augmentation to a human decision-maker conducting dynamic targeting decisions in a timecompressed environment. General description of optimization The historical starting point of optimization correlates with the advent of Operations Research 79 (OR) as a scientific approach generally attributed to military problem solving during World War II (Hillier and Lieberman 2021). According to Hillier and Lieberman, it was the “urgent need to allocate scarce resources to the various military operations and to the activities within

each operation in an effective manner” during the war, that led the British and later the U.S military having scientists explore what later became known as the OR approach and methods of optimization (2021, 1–2). Teams of scientists applied a scientific approach on pressing military problems and were instrumental in winning the Battle of Britain as well as in winning The Battle of the North Atlantic (Ibid. 2021, 2) The successes of OR during the war, found fertile ground shortly after the war. Two principal factors were to coincide and offer this: the industrial boom and the computer revolution. With the advent of the Simplex Method (Danzig 1963), devised by the mathematician George Dantzig in 1947 (Nash 2000, 1), OR expanded into organizational management and businesses along the axis of continued software development, personal computers and the “ability to 79 The term OR, is explained as “Research on (military) Operations” See Lundgren et. al ‘Optimization’, Lund:

Studentlitteratur AB 2010, 2. 188 | EXPLORING DECISION ADVANTAGES perform arithmetic calculations millions of times faster than a human being” (Hillier and Lieberman 2021, 2). Optimization denotes “the science of making the best decision or making the best possible decision” (Lundgren, Rönnqvist, and Värbrand 2010, 1), whereas ‘best’ indicate that there is a defined objective function to the problem, and ‘possible’ imply restrictions to the decision-making. A best solution is also indicating that there could be more solutions given the conditioning of the problem. It is a fundamental method in several fields beyond OR to include, Management Science, Mathematical and Computational Sciences. A prerequisite and a common denominator for all optimization models is that the objective and constraints can be expressed quantitatively in mathematical functions and relations (Lundgren, Rönnqvist, and Värbrand 2010, 2). The scope of optimization transcends both the

military, the academic and theoretical, deeply embedding itself into real-world practical applications within industry to include minimizing production time, satisfying customer demands in transport and logistics, and deciding the location of base stations and transmitters in telecommunication network design (Ibid. 2010, 3–7) Other fields such as health care apply optimization to manage resources sustainably, and financial planners use it to maximize returns while minimizing risk (Hillier and Lieberman 2021, 3). In the digital realm, algorithms optimize network flows, latencies and data routing to enhance internet and communication systems' performance (Yates et al. 2021) The rapidly increasing field of machine learning also leverages optimization to fine-tune models (Goodfellow, Bengio, and Courville 2016, 81), even if other areas such as deep learning which use neural networks still challenge the use of optimization models (2016, 317). The diverse and wide application and the

commercial interests driving it seem to be leading to further improvements of the method. Optimization methods begin with a clearly defined problem, often derived from complex real-world scenarios. To make these problems tractable, real-life complexities are distilled into a manageable and structured form – a model. The primary purpose of applying optimization techniques to model real-world problems is to derive valuable insights and identify feasible solutions that may be otherwise challenging to uncover through traditional analysis. Several motives for using models have been posited by OR-expert Williams (2013), including to recommend novel courses of action, revealing internal relationships and provide a greater understanding of the object being modelled (Williams 2013, 3). Similarly, other scholars reinforce this and argue that when using optimization models “we can simulate the real world and many scenarios can be tested to evaluate cause and effect with changed input data.”

(Lundgren, Rönnqvist, and Värbrand 2010, 11) The optimization process is an approach that includes some common phases as suggested by Lundgren et. al (2010, 8–11) Given that the real problem can be defined and that the problem is quantifiable, the first phase sets out to clearly EXPLORING DECISION ADVANT AGES | 189 identify the Real problem. This often involves determining which of the elements in the real problem are relevant, and which are of less importance. Some elements might be complex and difficult requiring them to be excluded, thereby forcing limitations and compromise to what the model can or will be able to calculate and consider. The result is a representation (model) of the real problem to be analyzed, and is referred to as the Simplified problem, which is a relevant and manageable account to the real problem. The simplified problem is then mathematically formulated as an Optimization model, which includes the definitions of the model’s objective function,

decision variables and constraints. The next phase is to apply a method to solve the model. The Solver can be a commercially available solver, or one that must be custom built for the specific application. To achieve an optimal solution for the model, verifications of the solution and validations of the accurate problem representation are required to evaluate the performance, after which the results can be obtained. A schematic overview of the complete process is visualized in figure 34 80. Figure 34 Schematic overview of the optimization process used in the experiment. The simple mathematical function (P) min f(x) s.t �� ∈ �� is a representation of the actual Optimization model. 80 The overview uses a mathematical function in Fig. 34 to represent the actual Optimization model’s mathematical formulations 190 | EXPLORING DECISION ADVANT AGES The process typically unfolds in several iterations, until the optimal solution is found. As indicated, the real problem may

have to be further simplified, or that modifications to the existing objective function, decision variables or constraints that bound the solution space are required to enable an optimal solution. As suggested by Robinson, an expert in modelling and simulation (M&S), ‘conceptualization’ of the real problem into a valid simplified problem is paramount and should comprise a non-software specific description of the purpose, objectives, inputs, outputs, content, assumptions and simplifications made (2017, 2741). This underlines the importance of iterations in an optimization process, as depicted in Fig. 34 Furthermore, to measure the results the solution must be verified, validated and feasible, or at least relevant to the intended use – this requires knowledge within the field of study. It is also worth mentioning that diverse types of models based on the same problem and data can have contrasting results, though at the same time extremely valuable to compare these. Such

potentially opposing results are implicitly indicators of the important relationship in-between the objective function, the decision variables and the constraints, and their independence of the data (Williams 2013, 4). Building optimization models requires defining three key functions: an objective function, decision variables, and constraints, where these functions can be expressed mathematically. This involves defining the objective function (one or more), with the most common expressions being either maximize or minimize 81. Two simple examples of a defined objective function are to ‘maximize the profit’ or ‘minimizing the cost.’ The objective function steers the approach to the problem and sets the aim. However, the objective function depends on decision variables and constraints that taken together bound (or narrows) the solution space. Decision variables are the parts within a problem setting that can be affected or controlled. Depending on the number of possible values

each variable can take, the more complex the model becomes, necessitating more computational power, or model simplifications. The constraints define the conditions or restrictions for the decision variables, that in turn, affect the objective function. Two basic types of constraints exist within optimization. Constraints are either considered as soft constraints, which indicate that they can be violated when the solver seeks the best or most feasible solution, they are considered to be hard constraints in which they must be satisfied (Hillier and Lieberman 2021, 265). Other considerations are the availability and reliability of the data used as inputs for the model. The complexity of the mathematical functions and the data will determine 81 For any given optimization problem, the objective function defines if the solution is to determine the maximum value of the objective function over the entire feasible region or the minimum. EXPLORING DECISION ADVANT AGES | 191 what

optimization method and computer support is required (Lundgren, Rönnqvist, and Värbrand 2010). Types of methods and problem classes There are different types of optimization methods and problem classes; however, it is beyond the scope of this thesis to describe all the various methods and classes. Instead, it focuses on the main types, the differences, and their application. Lundgren, Rönnqvist and Värnberg suggests that there are two main types of optimization methods: an exact method and a heuristic method (2010, 10). The exact method means searching for the optimal solution and verifying it. The heuristic method means establishing solutions of decent quality without an ability to estimate its deviation from any optimal solution. The choice of method is dependent on the nature of the problem and applied accordingly. Hillier and Lieberman argue that heuristic methods such as decision analysis (2021, 590), metaheuristics (2021, 634), and game theory (2021, 655), are designed and

applied to problems where the outcomes or inputs, are fraught with uncertainty. If, however, the problem has less uncertainties or can be simplified to reduce the uncertainties involved, then the problems can be solved using more exact methods (Hillier and Lieberman 2021, 655). Both main type of methods searches for optimality to provide a decisionmaker with a calculated conclusion Depending on how the problem is mathematically expressed it will obtain different type of classes (or categories) sharing the same characteristics. The most common are linear programming (LP), non-linear programming (NLP), and integer programming (IP). An integer is a whole number (not a fractional number) that can be positive, negative, or zero. A linear programming model demands linear expressions 82 , that is, it must have a linear objective function subject to linear constraints (Hillier and Lieberman 2021, 45). In other words, the linear expression is a mathematical formulation where the objective

function (ex. to minimize the cost) is the sought for solution which in turn depends on (or is a function of) the decision variables (ex: amount of expenses in a process), given a set of conditions (constraints) representing restrictions to the variables. Since all functions must be linear it could however become a “limitation for many practical problems” (Williams 2013, 21). However, their linearity also makes them easier to solve compared to nonlinear programming models. Furthermore, Hillier and Lieberman defines four assumptions to be implicit in a model formulation for linear programming: proportionality, additivity, divisibility and certainty (2021, 45–51). Accordingly, proportionality assumes that the contribution of each variable to the objective function is proportional to its value (2021, 45); additivity assumes that each variable contributes to the objective function independently (2021, 48); divisibility implies 82 A linear expression is a mathematical formulation that

represents a straight line (linear) when shown in a graph. 192 | EXPLORING DECISION ADVANTAGES that the decision variables can take on fractional variables (2021, 50). As discussed when presenting integer programming (IP), if variables must be integers, the problem becomes harder to solve. Lastly, certainty is the assumption that the parameters of objective function coefficients and the constraint coefficients are known and will not change (2021, 51). These assumptions support evaluating if linear programming applies to a given problem. If one or more assumptions for linear programming is violated, and the problem cannot fit the model, it may be possible to turn to either of the two models discussed next instead. The second class is nonlinear programming to which there are many diverse types of problems depending on the definitions of the objective function and the decision variables. Nonlinearities can arise in the objective function, when, for example experience is considered

within a problem formulation (experience will make production more efficient and reduce for instance cost in a nonlinear way). This will create various types of non-linear shapes to functions (Hillier and Lieberman 2021, 528). The last of the three types of classes is integer programming. In this type of problem area at least one variable is restricted to integer (or discrete) values. A number of situations require models to apply integer programming, for example deciding how many sensors to allocate or weapons to use. These variables can only take a set of discrete values. It means one cannot use two and three quarters (275) sensors or weapons. Another application of integer variables is whenever the variables are restricted to two values, 0 or 1. Most practical integer programming models use the binary variables 0-1 to represent ‘yes or no’ decisions 83 (Williams 2013, 155). The use of binary variables can help to reduce the complexity of the problem leading to faster and more

accurate solutions. However, adding more binary variables can increase the number of potential solutions exponentially, leading to complexities and no way of solving the problem in reasonable time due to computational difficulty and an increased requirement of computer memory (Williams 2013, 156). Despite the differences in how the problem is formulated, optimization is intended to inform and augment decision-making. One of the most common application involves allocating (or assigning) limited resources among competing activities in a best possible way (Hillier and Lieberman 2021, 32, 338). 83 As a simple example: if we are to create a work schedule in which we must ensure that no employee double-booked, then using a binary variable helps simplify the scheduling decision process – either the employee is working or not. EXPLORING DECISION ADVANT AGES | 193 The military problem: A weapon to target assignment problem The weapon to target assignment (WTA) problem is of military

importance because it computes an optimal solution to assignments of n weapon 84 to n targets that meets the objective(s), often defined as maximizing the expected damage to a target, or minimizing the expected target survivability (Lu and Chen 2021, 1). A WTA problem relates both to the transportation problem (Williams 2013, 82), which strives to obtain the minimum cost flow and the assignment problem (Ibid. 2013, 87), which refers to the problem of assigning n people to n jobs so as to maximize some overall level of competence. Both problems are defined as network models with historical roots in the Hungarian method further exploited and described by mathematician Harold Kuhn, in 1955 (H. W Kuhn 1955, 83–97), and originally introduced as a WTA problem by OR-researcher Alan Manne (1958). The WTA problem is a combinatorial optimization problem that involves assigning a number of weapons to engage a set of targets that optimizes mission effectiveness. As the number of weapons are

finite and limited their employment (or assignment) to various targets can be optimized. In a military targeting setting, the closest type of class would be an integer programming model and relating to an assignment problem where various own resources (weapons) are being assigned to specific tasks (missions) against emerging targets of an adversary, subject to relevant constraints, to include ensuring that each resource is assigned to only one target. This problem has been formulated and solved in diverse ways for decades. According to a recent survey on WTA, the majority of literature on the topic focuses on the defensive perspective (how to defend against incoming threats), as opposed to a more targeting-related offensive perspective where a force tries to optimize its targeting process by optimizing its employment of own resources in targeting an adversary’s resources (Kline, Ahner, and Hill 2019, 226). Their survey covers the evolution of various approaches to the common WTA

problem, which they argue involves two distinct categories of WTAs: static (SWTA) and dynamic (DWTA) with the main difference between the two residing in that the latter incorporating more than one stage, and in doing so implicitly includes time as a dimension (2019, 226). Both categories still epitomize the quest to optimally match elements of two sets (weapons and targets) based on given criteria, ensuring that an element from one set is matched to only one element in the other set. 84 In this context n is a mathematical symbol representing different weapons and different targets (being the two main decision variables) in a WTA problem setting (and subsequent mathematical formulation). 194 | EXPLORING DECISION ADVANT AGES The real problem The framing of the real problem is briefly presented in a condensed form, before two additional attributes (speed and complexities) are discussed together with a narrative to further illustrate the real problem, prior to building a simplified

representation of the problem. Joint targeting involves complex processes, subtasks, and activities. So too, does the more general framework of Boyd’s OODA loop. However, the review of literature in Chapter 2 reveals that military organizations requires decision support systems specifically designed to augment faster decision-making (Budning, Wilner, and Cote 2021; J. Johnson 2022a) This is further developed in Chapter 3, Section - Joint Targeting, discussing the strive to enable ‘kill chains’ (B. Johnson et al 2023, 155) or even ‘kill webs’ (Penney 2023, 1) to make targeting more efficient and effective. The explicit research review on optimization algorithms gives the same impression (Hocaoğlu 2019; Park and Choi 2023). A newly published MIT technology article (‘War Machines’) states that the US Army has managed, “to shorten its own tactical 20-minute targeting cycle to 20 seconds” (Holland Michel 2023, 3). The Ukrainian army uses a program, GIS Arta, that

supports command and control by pairing Russian targets on the battlefield with the closest Ukrainian artillery unit which significantly reduces the time to less than a minute (Kobzan 2022, 7). Beyond these arguments there is the alleged efforts (Davies, Mckernan, and Sabbagh 2023) claiming that Israeli Defence Forces (IDF) have built AI systems to augment targeting-related decisions in an effort to make the targeting process more effective and efficient. These are all arguments for faster decision-making and a speedier targeting process. The speed of decision-making The speed in which decisions are turned into action is highlighted and amplified in the synthesis (Chapter 2) and is considered inherent in Boyd’s OODA loop. Although some of the examples illustrated in the previous section are related to army targeting (tactical level) and therefore should be essentially swifter than the higher command level (operational level), the fundamental factor that amalgamate all of the

above-mentioned examples is speed. Somehow, the joint targeting, or for that matter, targeting in general must improve its ability to execute targeting faster. In turn this necessitates the abilities to understand – decide – act 85, or using Boyd’s OODA loop, the abilities to observe – orient – decide –act. The former is arguably a simplification of the latter implicitly incorporating the input of data and information prior to being able to understand anything. The efforts to reduce the time navigating through a dynamic targeting process starting with the detection of an emergent 85 Used by some researchers, for example (Penney 2023)to explain the kill chain. In this thesis, it is used to explain the inherent work of non-human intelligent agents. They are arguably all interpretations or simplifications of Boyd’s OODA loop EXPLORING DECISION ADVANT AGES | 195 target to engaging it, is analogous to Boyd’s OODA loop from observation via orientation and decision to

action. It underlines not only the relevance of this experiment, but it also positions the theoretical framework of Boyd’s OODA loop at the center. To further illustrate the real problem in a fictive context, consider the following imaginary narrative: Being delegated the authority to run the day-to day management of the joint targeting efforts under the JFC, you have some twenty or so people working close to you monitoring the work, and another dozen at each of the subordinate commands. All the weapons that could be utilized for target engagement are updated frequently in a database, and the different targets management lists and target folders are in another database. In a specific document, the current targeting directive with its objectives, desired effects, and guidelines to support the decision-making process are available. Now, though some targeting activities are managed deliberately, and therefore preplanned days ahead, other targets of opportunity may suddenly emerge and

pose challenges for the joint force or a threat against it. These needs to be dealt with as soon as these targets are reported. These inputs (reports) from diverse intelligence sources are available in the form of data and information. The problem then becomes - by what means (weapons) should each specific target be engaged? Or in optimization terms, how can we maximize the probability of kill, given a number of business rules (constraints) and policies (multi-objective criteria)? This narrative illustrates a simple problem of matching weapons to targets. However, there are more to consider than this, making targeting a complex undertaking. The complexities in targeting The complexities in targeting are challenging for any targeting enterprise, especially when dealing with targeting under dynamic conditions. General concerns that a targeting enterprise must account for prior to decision-making include: the availability and status of the weapon and its potential platform; its

abilities to effectively be able to engage; the operational range of own resources; individual target characteristics; the location and protection level of the targets; the surrounding area for collateral damage estimations; if the targets are stationary or mobile; a targets precedence over other targets considered for engagement; the current decision policy and potential constraints; the desired effect on the target as well as resources to support each engagement with terminal guidance; preplanning of post engagement assessments, and preparing to perform re-attacks if required. These are all relevant considerations. However, this is not a complete outline of all the factors accounted for. For a more complete outline of all the considerations and steps done, see US Doctrine JP 3-60. Yet, it reveals that targeting decisions are a complex undertaking. 196 | EXPLORING DECISION ADVANT AGES The experiment – a concrete example of the military problem This section addresses the

simplification of the real problem, the experimental setup including how the model is built and evaluated following the process displayed in Figure 6. It outlines the data used and the processing steps following the generic sequences of the optimization process seen in Figure 6. The last part explains how the model’s performance is evaluated, including its decision-making capabilities. The study employs a multi-stage and multi-criteria model aimed at replicating a dynamic targeting environment and linearizing the probability of kill (Pkill) function within, to enable an integer linear programming model (ILP) (Hillier and Lieberman 2021, 460). The term probability of kill, or kill probability, is a commonly used term for WTA-problems and refer to the likelihood of destroying a target in an attack (see for instance (Metler et al. 1990; Kline, Ahner, and Hill 2019; Chang et al 2023b)) The study therefore presents a new way of modelling WTA problems, and a comprehensive approach where

AI (the optimization model) augments a joint force’s timely targeting decisions by optimizing the use of joint weapons. The objective of the experiment is to develop an efficient multi-criteria rule-based AI-application to solve a dynamic weapon to target assignment (DWTA) problem. The problem (a scenario or mission) is considered solved if all targets presented to the model are engaged and the desired effect of destroy is accomplished. The problems and solutions are, however, subject to several multi-objective criteria and constraints. Using optimization is, by reasons brought forward in the next section, is a feasible approach to meet the objective of this experiment. It comprises the following aspects: • • • Design of a DWTA model considering a variety of emergent targets of concern for a joint force, and realistic operational and technical constraints on the attack options restraining the possible assignments. A multi-stage approach based on a linearization of the

probability of kill function (Pkill) defined to either meet the threshold of 0.9, or to engage in a re-attack in a second stage. Linearization ensures the quality of each solution every time, compared to a heuristic approach where there is no guarantee of solution quality. An optimal solution is guided by the multi-objective function where five (5) different arguments are prioritized using a lexicographic approach. Each of the five arguments is perceived as part of a targeting decision policy. A change to a targeting decision policy requires changing the objective arguments to replicate the decision policy. EXPLORING DECISION ADVANT AGES | 197 • The problem can be solved in a few seconds using a single workstation (laptop), thereby allowing time for the human decision-making process (OODA-loop) to consider the options computed by the optimization model. The performance index for evaluating the assignments is determined by four criteria: (1) the ability to recommend feasible

attack options of all targets provided by the decision-maker; (2) the ability to handle prioritizations; (3) the ability to perform re-attacks if the threshold for probability of kill (Pkill) is less than 0.9, and lastly; (4) the ability to minimize the cost of the overall target engagement. Beyond these criteria, the model is evaluated by using sensitivity analysis, CPLEX 86 Optimizer’s internal validation protocols, and in-person verifications of solutions. Moreover, the experiment is deliberately framed as time-sensitive and is more related to the doctrinally defined dynamic method than the deliberate method (see Chapter 3). The dynamic method is used to address targets that unexpectedly emerge (emergent targets) in the battlespace. Once identified, these targets would require fast, accurate, and reliable decision-making, where minutes and even seconds could change the situation to a successful outcome. If dealt with successfully, these emerging targets could be perceived as

opportunities rather than threats. Consequently, military organizations are likely to strive for speed - AImodels that recommend solutions rapidly and can explain why it arrived at its decision has significant benefits. That is also what Model 2 is built and intended for The rationale for the model can therefore be motivated as a joint force involved in dynamic targeting activities during a military operation. Furthermore, the targeting activities would be subject to several considerations intended to reflect a predefined decision-policy that reflect a commander’s current targeting policy. This includes the precedence between targets so that the joint force targeting priorities can herald even when the operational pace increases. This notion does not exclude other target engagements with targets that pose a threat to the force but does exclude combat engagements. The level of intelligence and foreknowledge of an adversary’s fielded systems will affect the ability to augment

automated, and predefined weaponeering 87 options that considered a target’s function, characteristics, vulnerabilities, and level of protection (US Air Force 2021). Furthermore, the weapons must be effective against the specific target to create the desired effect. The matching of weapon-to-target is constrained by inherent capacities, the protection level of the targets, and is subject to probability in the execution of a target engagement. High priority targets may require weapons with high accuracy (precision-guided munitions) or other distinctive features. Such weapons are more expensive and therefore fewer in numbers. Alternatively, other less expensive weapons may have the same effect A 86 An IBM software package to manage and solve optimization problems. 87 Weaponeering is defined as “the process of determining the quantity of a specific type of kinetic or non-kinetic means required to create a desired effect on a given target” See US Air Force ‘Air Force Doctrine

Publication 3-60, Targeting’, US Air Force, 2021, 50. 198 | EXPLORING DECISION ADVANT AGES suitable example is a target that is currently static. Its static state enables more, and likely cheaper weapon alternatives, than if the target is moving. Moreover, emerging targets are associated with greater uncertainties, owing to a lack of intelligence on these targets. The experiment assumes that, from a ‘real problem perspective’, all prioritized emerging targets require to be targeted. This notion allows us to assume that a re-attack to reach the desired effect intended is necessitated. Simplified problem Simplifying a problem makes it manageable for modelling and simulation purposes. A dynamic targeting problem can be fixed in its form and then further simplified by using chain-analogy. The chain can be assumed to start with detecting a target and end with engaging it. This also correlates largely with the OODA loop An observation is here seen as a detection, ending with

acting which is equivalent to engagement. Both the targeting cycle and the OODA loop involves subsequent feedback or assessments to support the understanding of the results. 88 As such, the chain involves several carefully orchestrated processes and tasks 89 to complete this loop. In reference to the description of the real problem the following actions have been taken to both narrow the scope and to make the problem simpler: Sensors are excluded from the problem formulation. Assigning sensors to accompany or complement effectors (weapons) is neither complicated, new (when relating to previous research dealing with this (Quttineh et al. 2013; Bogdanowicz 2009)), or appropriate within the classical weapon to target assignment problem. The optimization model (OM) only considers physical targets that are on the ground. All physical target coordinates to include the specific joint desired point of impact (JDPI) for the specific target are assumed to meet the requirements and policy for

target coordinate mensuration(TCM) 90 . This assumption involves to generate precise three-dimensional coordinate data (latitude, longitude, and elevation) and associated uncertainty estimates (US Chairman Joint Chief of Staff 2022, 15). Although joint targeting accounts for all types of weapons that can create an effect on a target considered for prosecution, the model is deliberately limited to the use of weapons such as bombs, missiles, and artillery munition. No cyber weapons, information activities, directed energy weapons (DEWs), including lasers and highpowered microwave (HPM) systems are therefore included, as they present a 88 As contemporary targeting is becoming more digitalized (see Chapter 2), it could be stated that the chain begins and ends with sensors. 89 See for instance US JP 3-60: advanced target development, which is a detailed process in the final stages of the joint targeting cycle, which comprises detailed target descriptions, precise target point

mensuration, and cautious collateral damage estimations. 90 Target mensuration refers to the generation of precise and accurate geographic measurements of coordinates. It is commonly referred to as precise point mensuration. The hardware and software for target mensuration along with procedures and policies are available and existing today. See Chapter 3 and US DoD ‘CJCSI 350501E’, US DoD, 2022 EXPLORING DECISION ADVANT AGES | 199 greater challenge to assess. Furthermore, the weapons 91 used by the model are chosen to resemble a reasonably credible joint force. The collection of weapons is selected purposefully to give the model assorted options to nominate. This means that there are always two or more alternatives for the optimization model to decide between. The assortment of targets is made based on the author’s own experience (from serving in different joint force settings) of how diverse military targets can be perceived as having precedence. The most important

aspect in the experimentation is not any predominant discourse on priorities in-between targets, but the very fact that there is a difference, so that the model incorporates, and accounts for three levels of priority. These levels are commonly found within the western targeting doctrines to help differentiate precedency. Assuming that friction and uncertainties are always present in all wars or conflicts, the simplifications made in this respect include one aspect of uncertainty to prove that it can be applied and utilized by the Solver in a way that makes sense. In addition, the study shows that this probability of kill (Pkill) function can be linearized. No collateral damage estimation 92 (CDE) levels are applied (see Chapter 3). All targets are presumed to be assessed as CDE level 1, making all weapons used in the model feasible choices. Time is not accounted for in the optimization model (OM). Most of the weapons (or effectors) used travel at 0.7-1 km/second The OM only considers

if the effectors are within range for target engagement or if they are not within range. Some highly mobile effectors, such as aerial drones and aircraft have an extended range since these platforms can move within its respective weapon’s effective range. Further, the limitations of different effectors from wind, visibility or temperature are disregarded, considering that these variables were assessed by the SMEs as minor concerns. Finally, simplifications have been made to targeting effects. The real problem includes different desired effects intended to support specific military objectives. Moreover, if a joint force also has non-kinetic capabilities, for example cyber weapons, then other effects may be applicable such as detect and deceive (Grant 2023). Typically, the effect of targeting refers to one of the following: deny, degrade, disrupt, or destroy. In this experiment, only one effect is considered – destroy 93 Destroy means that damage done to the function is permanent

(NSO 2021). The 91 Weapons are used interchangeably with effectors. Both terms take into account the platform (artillery, aircraft, drone) and the munition (shells, bombs, missiles). 92 CDE refers to a methodology used in support of targeting decision-making, in which collateral damage (civilian casualties and damage to civilian objects) is assessed prior to target engagements. It is a deliberate and mandatory step that also provides the necessary base for a measured approach to proportionality (one of the cornerstones in the Law of War). The methodology is described in detail in the US ‘CJCSI 3160.01D’, US DoD, 2018 93 Destroy – the term refers to means permanent damage done to the function of the target. The function's operation is permanently impaired including all facets of the function's operation. See NATO ‘Allied Joint Publication for joint targeting’, NSO, 2021 200 | EXPLORING DECISION ADVANT AGES other three effects are concerned with a lesser degree

of damage to the target’s function or can even be made to deny the target’s ability to communicate. However, the choice is motivated by its clarity and the fact that it requires more resources than the other effects do. Data set The data set for the study was compiled in three steps after a pre-study presentation for a small group of subject matter experts (SMEs) (19 Sept 2022). The SMEs were active personnel within the Swedish Armed Forces representing the Air Force, Army, and Navy services. They were selected to provide feedback on the experimental set up based on their respective domain knowledge and their perceptions of the current practical targeting challenges between the joint level and the component levels. The admission on their part for the value of such an experimental setting became an additional ‘quasi-external’ confirmation for the study. Some of these SMEs were supportive in validating the internal consistencies within the weapons and target tables (data

collection) to include confirming weapons data and complementing the data set with general features and characteristics for some of the more specific weapons used in the data set. The data set consists of three different databases. The first is the effectors database that contains all available weapons and the various weapon platforms. The second is the target database structuring and defining all the targets and their individual characteristics. Finally, the third database is the compatibility database, which matches compatible targets and effectors in diverse ways termed attack options to reflect the approach including the type of effector, the type of weapon, and the number of units of weapons. The effector data set used consists of 13 diverse types of weapons (the term effectors is used interchangeably with weapons) used by 6 different types of platforms as described in Figure 35. Figure 35:The table shows the effectors (weapons) used in the data set and their corresponding

platforms. The unit of min/max distances to targets are kilometers EXPLORING DECISION ADVANT AGES | 201 The table in Figure 35 defines the identification of each effector used (Effector ID), that is, the platform that operates a certain weapon type. In the simplified problem setting, the data set still contains some internal factors that can be manipulated (modified) to trigger the optimization model’s (OM) variables to select alternative effectors. The total amount of a specific weapon type can be set to 0 (zero), which informs the OM that that weapon type is inaccessible. Additionally, the simplifications still include some notable features. The effective range of an effector relates to its relative distance (in kilometers) to each emergent target as presented to the OM. If the targets are beyond the effector’s range (for example, if artillery is out of range) the OM will recognize this and choose another effector that is within range. Aerial platforms are a special case of

effectors in the data set, since theoretically, they can rapidly close their relative distance and subsequently engage their carry-on weapons. Time, to include travelling time for effectors are not an integrated factor in the OM. Instead, the OM is given a binary input (yes/no) as to whether the effector has extended engagement range. All weapons and aircrafts (excluding drones) travel at 0.7-1 km/second, which means that target engagement duration tends to be a matter of minutes from decision to action, making the preceding phases (faster understanding and reaching a decision) the most valuable to improve. Additionally, the effectors’ individual ability to effectively engage moving targets is considered. To engage a moving target requires that advanced capabilities be integrated into the targeting enterprise. It includes tracking capabilities or sensors and capabilities beyond just a GPS and an Inertial Navigation System. The additional abilities comprise integrated data link

allowing for post-launch tracking and control of the weapon itself, multi-mode seeker (radar, infrared, laser) to provide engagement abilities in adverse weather conditions (FY18 Air Force Programs SDB II US Department of Defense 2018, 201). In the table, these abilities are integrated in the effectors that have ‘Moving’ defined in the column ‘Target location,’ indicating the effective engagement of moving and maneuvering targets to the OM. To enable cost and benefit calculations and comparisons between scenarios, the last column in the table defines the cost of each weapon per unit in US dollar. The costs have been retrieved primarily using the manufacturers’ web pages. An additional table was also developed, defining not only the total amount of each type of weapon (in units), but also the number of platforms and their respective distribution over a defined geographical area. It was never used in the different scenarios, mainly because the results were abundant as they

were. Another portion of a WTA problem concerns the target. 16 different targets were included in the data set as seen in Figure 36. 202 | EXPLORING DECISION ADVANT AGES Figure 36: The table defines the 16 types of targets used in the experiment. The table in Figure 36 shows the target identification number (Target ID) along with related key factors or characteristics of each target that will inform the optimization model (OM). From left to right, the second column indicates each target’s significance, meaning, within a joint force, to whom the target is important and what are the implications. Three levels of significance are used: Time-sensitive target (TST), high-value target (HVT), and component critical target (CCT) (see Chapter 3). These have a close (but not mandatory) relation to the next column ‘Priority,’ defined as numbers ranging between 1-3, and where 1 refers to the highest priority ranking. The column ‘Category’ uses the naming convention of NATO target

set defined in Allied Joint Doctrine for joint targeting (2021) and the subcategory ‘Target type’. The different targets were selected based on three criteria: providing an internal capability to match the targets with one or more effectors, to pose different threats to a joint force and require target prioritizations, and diverse complexity levels. The latter criterium is further defined in the two subsequent columns, where these are created as derivatives of the specific target type. For example, a ‘Bridge Large’ 94 (T14) comprises three target elements (important to the function of the target - a structure used to get across parts of the terrain in the operational environment (OE)), where all three are considered as critical/vital elements . Logically, the smaller the bridge the fewer target elements and critical elements exist. 94 Although not of vital importance to the study, the differentiation was based on the following: Large >100m w >3pillars (>5 JDPI), Medium

<70m w 2 pillars (3-4*JDPI), and Small < 30meter in length (1JDPI). JDPI is the abbreviation for joint desired point of impact. EXPLORING DECISION ADVANT AGES | 203 If we examine target ‘T3’, we find that the medium range surface -to-air missile system comprises 10 target elements, yet the table only considers 5 of them as critical elements (vital to the function). Each of these cases, in which a difference between the two elements exists, has prompted a systematic analysis. In terms of ‘T3’, the analysis considered the composition of a complex yet typical medium-range SAM system (9K317 "Buk-M3), which includes 1* 9C510M combat control station (PBU), 1* 9C18M3 target detection radar station, 6 9A317M autonomous selfpropelled firing units (SPS), and 2 transport and charging machines (TZM) 9T243M. This composition constitutes 10 target elements, of which the control station, the radar station and three firing units are considered critical, to fulfill the desired

effect ‘destroy’. The last column represents the targets protection level. Three levels are considered: vulnerable, neutral, or protected, and these can be manipulated to offer comparison between different solutions. The uncertainty of a target’s specific level of protection is seen as a significant factor to consider. Protection could be due to the use of jammers or decoys as well as, for instance, high level of readiness. The protection level was introduced to the data set to have correlations towards the computation of the probability of kill function (Pkill) by the OM. In each scenario, the OM’s modfile 95 is ‘unaware’ of the level of protection that has been manipulated in the specific dat.file 96 The given input is the results of the first stage of attack, in which the manipulation mimics real world uncertainties, and the OM manages the problem of not reaching the level of Pkill by re-attacking in a second stage. The compatibility table contains all feasible attack

options (targeting assignments), where each individual set of attack option a is an element of all feasible attack options. The complete table includes all feasible weaponeering options, including the preferred options for each weapon to target matching. The effectiveness of an effector against different targets is assessed based on the specific characteristics of both the effector and each individual target. The compatibility table was validated by the SMEs, to the extent of their expertise and ability to provide unclassified information. In some instances, where no external reference was available, the author made some approximations to complete the data set. An example of the table discussed is presented in Figure 37. 95 The mod.file is the model file that includes the algorithms (instructions) that solves the optimization problem 96 The dat.file refers to the different data sets that the model (modfile) turns to as input data for its reasoning 204 | EXPLORING DECISION ADVANT AGES

Figure 37: The table shows a portion of the Compatibility table that was used as the main input to the Optimization model (OM). Please note that the last two columns are manipulated to fit a specific set up/ scenario. EXPLORING DECISION ADVANT AGES | 205 The table depicted in Figure 37 represents a portion of the complete compatibility table used for the model as one of its input data sets. It shows the options available to execute a target engagement if a target were to emerge and be observed on the battlefield. It gives the optimization model (OM) options to consider given its predefined decision policy (subject to the specified objective function(s), decision variable(s) and constraints) during the orientation and prior to its decision on actions that the OM recommends when augmenting a human decision-maker. In this particular example, as visualized in Fig. 37, six distinct types of effectors (E2,E5,E8,E10,E12,E13) are compatible with three different targets (T9,T10,T11),

resulting in a number of attack options (A163-A192), and subject to each effectors’ availability for the decision-maker. The sum of all attack options for the OM within the compatibility table used in this experiment is 332, considering its 16 target types and 13 type of effectors. Each available attack option has its individual characteristics, some of which are susceptible to manipulation depending on the specific scenario. Starting from left to right (Fig. 37), the first column defines the specific attack option’s ID (‘Axxx’) that offers explainability and transparency to each individual decision (output) made, thereby enabling a logical trace back to the data input. The following columns reference each target with an ID, category, and type, after which a matching is made to a specific effector (‘Effector ID’), and its corresponding type of weapon (Effector Type). Next are three columns stating the number of weapons used pending the number of stages, after which the

corresponding Pkill follows. Column L reflects whether or not this option is to be a multi-stage attack, and the last two columns (M,N) are susceptible to manipulation to define if the target is within range of the specific effector or not. Furthermore, the compatibility table reflects the following considerations: A. B. The second stage in a multi-stage attack option implies that a decision has been made to re-attack in order to create the desired effect (destroy). However, this process is simulated (automated) as there are no external sensory inputs of any actual effects from the first attack. § by the OM: if the target’s protection level was set as vulnerable, neutral, or as protected. The number of weapons required against a specific target is determined by achieving a minimum probability of kill (Pkill) of 0.9, which indicates the target is destroyed. To ensure a realistic relationship between the recommended number of weapons and any reduced quantity, an adjustment factor,

(10,05*(0,9-0,4275)) is applied to provide sufficient differentiation between related attack options. Additionally, if a target is classified as ‘self-protected’ (ie, equipped with countermeasures), for example, a Surface-to-Air Missile System (SAMS) in categories T3-5 or a potentially protected large structure like a bridge 206 | EXPLORING DECISION ADVANT AGES (T14), then one extra weapon is added to the required count to ensure the desired Pkill of 0.9 For instance, if 3 weapons are calculated to be sufficient for a target without countermeasures, then a protected target would require 4 weapons (3 + 1). This accounts for realistic adversary countermeasures and provides a more robust engagement plan. For example, destroying five critical elements with a weapon designed for precision, such as the AGM-88E, would require six missiles instead of five, to counteract the impact of potential defenses. Note: The concept of Pkill here does not apply to targets classified as

‘Vulnerable’ or ‘Protected’ in a multi-stage engagement, as these are covered separately in section D. C. D. E. F. G. Only effectors that have the capability to engage moving targets can be considered by the OM. Hence, their Pkill is not reduced by the fact that a target is moving. In a re-attack (multi-stage), targets defined as vulnerable increase the Pkill value by 0.15, whilst protected targets decrease the Pkill value by 015 Each scenario predefines the state of protection for each target determined in the preprocessing of the specific scenario. This makes the model deterministic, rather than stochastic. Two targets (T6, T16) are considered not to be protected due to their specific attributes. Most precision-guided munitions (PGM) have a declared circular error radius (CER) of 1m. Weapon may have Airburst/ Submunition for specific targets, making them the preferred choice. In a second stage, the Optimization Model (OM) adjusts based on the results from the first

stage. A re-attack the purpose of the second stage would traditionally be conducted after a physical battle damage assessment (BDA) determines that the initial attack did not achieve the desired effects, resulting in a re-attack recommendation (RR). In the OM, however, this RR is simulated based on outcomes from the first stage rather than using actual BDA or RR inputs. The number of weapons assigned in the first stage represents a variation of the actual impacts observed after an initial BDA. This means that no reoptimization is conducted It is an action following the result of the previous stage for each individual target. This accounts for unforeseen factors such as weapon malfunctions, targets being obscured in the last moments of an attack, or operational errors. These variations are reflected in the different rows representing each effector engagement option. EXPLORING DECISION ADVANT AGES | 207 Mathematical model of the decision environment A mixed integer linear

programming model (MILP) was used to operationalize the conceptualization of the simplified problem setting. Since the simplified problem contained some variables that were constrained to be integers while others were intentionally chosen to be continuous, the model’s construction became a mixed integer model. This enabled flexibility (Williams 2013, 207) in combining discrete (e.g targets, effectors, units of weapons) and continuous elements (probability of kill and costs) to solve the complex optimization problems at hand, which was perceived as essential. However, as suggested by Williams (2013, 208), one might expect the solution time to grow exponentially with the number of 0-1 variables since each integer indicate 2n (power of two), which means the number of solutions is doubling with each n (n being the number of integers). 97 This pattern is referred to as the exponential growth of the difficulty of the problem(Hillier and Lieberman 2021, 473). On the other hand, and opposite

to an linear programming model, expanding the number of constraints can make an integer model easier to solve. (Williams 2013, 211). The optimization model uses three different data sets with respective symbols: Targets (��) where each individual target �� is an element of ��, Effectors (��) where each individual effector �� is an element of ��, and Compatibility set ( �� ) which contains all feasible attack options (targeting assignments), and where each individual set of attack option �� is an element of all feasible attack options �� . As not all effectors ( �� ) can attack every target , a compatibility set (��) is introduced to make the model more effective and compact. The compatibility set consists of the compatible (or feasible) effector-target combinations, rather than considering all possible �� × ���� × �� combinations. Consequently, the decision problem consists in optimizing the best sets of attack options

given the targets, the decision parameters and constraints. The following parameters (which are provided as input data) are used to further support formulating the problem: ���� ∈ {1,2} is the priority of target �� ∈ ��, ���� ∈ {��������������������������������������, ��������������, } is the protection level of target �� ∈ ��, ���� ∈ �� is the target attacked via attack option �� ∈ �� ���� is the kill probability via attack option �� ∈ �� ���� is the number of weapons used in attack option �� ∈ �� ���� ∈ �� is the effector (weapon used in attack option �� ∈ �� 97 ‘Power of two,’ or 2n, in which means that 2 is multiplied by itself n times giving 21= 2210=1024. 208 | EXPLORING DECISION ADVANT AGES ���� ���������� is an

indicator equal to “1” if attack option �� ∈ �� is multi-stage; otherwise “0” ��1�� ���������� equals “1” if target in attack option �� ∈ �� is in range of stage 1; otherwise “0” ��2�� ���������� equals “1” if target in attack option �� ∈ �� is in range of stage 2; otherwise “0” ������ ���������� equals “1” if effector �� ∈ �� supports attacks in range of extended engagement; otherwise “0” ���� is the available quantity of effector �� ∈ �� (weapon), ���� is the unit cost for each effector �� ∈ �� (weapon), �� is the budget constraint in given scenario. ���� is the realized probability of kill of target �� ∈ �� based on the specific attack option (weapon assignment decisions), and represents the total probability, considering all the weapons associated with that attack

option ���� equals “1” if attacking option �� ∈ �� is selected in the solution; otherwise, “0” ���� equals “1” if target �� ∈ �� is killed; otherwise, “0”. A target is said to be is killed if the probability of kill is above 90%. If the realized probability of kill of target �� ∈ �� exceeds the threshold then target is defined as killed (���� is equal to 1), otherwise target �� ∈ �� is defined as not killed (���� is equal to 0). ���� equals “1” if effector �� ∈ �� is utilized in any attack; otherwise, “0” ���� is the total number of effector �� ∈ �� utilized in attacks The model has the following constraints: (1) At most one compatible effector attack can be assigned to each target. EXPLORING DECISION ADVANT AGES | 209 � ��∈��∶���� =�� ���� ≤ 1 ∀�� ∈ �� (2) Effectors used cannot exceed associated

available quantities. ���� ≤ ���� ∀�� ∈ �� To support this constraint a complementary condition is required to keep track of the number of weapons utilized. Calculate number of weapons utilized 98. � ��∈��:���� =�� ���� ∗ ���� = ���� ∀�� ∈ �� (3) The total cost of the operation/mission cannot exceed the associated budget. � ���� ���� ≤ b ��∈�� (4) Define realized kill probability of each target. � ��∈��:���� =�� ���� ∗ ���� = ���� 98 The quantity is decided based on the selected sets of attack options which is an output of the model. 210 | EXPLORING DECISION ADVANT AGES ∀�� ∈ �� (5) Distance constraints #1: if attack is single stage and if target is out of range of Stage One, and effector is not drone/aircraft, then that attacking option is not possible. ���� = 0 ∀�� ∈

��, �� ∈ �� ∶ ���� = ��, ���� ���������� = "0", s1r flaga ="0", eer flage ="0" (6) Distance constraints #2: if attack is multi-stage and if target is out of the range of Stage Two, and effector is not drone/aircraft, then that attacking option is not possible. ���� = 0 ∀�� ∈ ��, �� ∈ �� ∶ ���� = ��, ���� ���������� = "1", s2r flaga ="0", eer flage ="0" EXPLORING DECISION ADVANT AGES | 211 Description of the mathematical optimization problem The model has five arguments (or criteria) to its objective function. It uses a lexicographic approach to solve the problem wherein each argument is assigned a strict priority ranking. In this hierarchy, the first criterium holds the highest importance, followed sequentially by the second criterium, then the third, and so forth. Each criterium is optimized

in order of priority, with higher-ranked criteria taking precedence over those ranked lower: (1) Maximize the total number of “1st priority” targets killed ���������������� (2) Maximize the total number of “2nd priority” targets killed ���������������� (3) Maximize the total number of targets killed � ���� � ���� ��∈��∶���� =1 ��∈��∶���� =2 ���������������� � ���� ��∈�� (4) Minimize Total Cost ���������������� � ���� ���� ��∈�� (5) Maximize Total Number of Different Effectors Utilized ���������������� � ���� ��∈�� 212 | EXPLORING DECISION ADVANT AGES Computational implementation and solution The model was computationally implemented by using a modelling language provided by

the solver (CPLEX) to define the problem's variables, objective functions, and constraints. Coding was conducted in a language termed ‘opl’ The problem was solved using IBM ILOG CPLEX Optimizer, edition 22.11 This software was selected on the basis of availability from the software considered to be exact and appropriate 99 solvers for a mixed integer programming problem (MIP). Four different versions were constructed (Version 1 to 4), each adding more considerations and complexity towards the last version (Model 2). All models were pre-processed and debugged using CPLEX debugging capability (Watson and Cacioppi 2014, 19; IBM Corporation 2013, 92) to ensure that the coding was correctly written. By progressing in model complexity from Version 1 to 4, attention could be paid to the specific additions, rather than the complete model. It also made sensitivity analysis easier. In addition, each version can also be understood as solving different, though related problems. For

example, if a user has a single stage problem without additional constraints then version 1 can be employed. Version 1 considers a basic problem setting with no probability of kill (Pkill), and therefore all targets are set as ‘Neutral’. Moreover, it only accounts for a single stage and one criterium - to maximize total number of targets that are killed (creating the desired effect of destroy). Version 2 considers multi-stages and introduces probabilities relating to if a target is ‘Vulnerable’, ‘Neutral’, or ‘Protected’. Version 3 involves multi-stage and multi-criteria, where the latter follows a specific multi-criteria policy of priority using a lexicographical function (staticLex) within the objective function. A brief explanation is given to staticLex , which was incorporated in both Version 3 and 4. The multi-criteria policy can be perceived as a defined decision policy outlined within a Commander’s targeting directions and guidance. It can therefore be

altered and changed accordingly, depending on the preferences, rules or strategies that follow in the targeting directions and guidance. The first criterium refers in Version 3 to maximize the total number of targets that are killed (criterium #3). The second criterium states that the model should minimize the total cost, hence choosing the most cost-effective option available to the model out of the roughly 300+ combinations. Finally, the third criterium is set to maximize the total number of effectors used in each scenario. The lexicographical function (staticLex) ensures that the model searches for an optimal solution in three iterations, one for each criterium. After the model finds an optimal solution for the first criterium, it runs the sequences again adding the second criterium to its 99 Hillier and Lieberman define a few ‘elite’ solvers, to include CPLEX, GUROBI and CoinMP in chapter 12 (Integer programming) of their book ‘Introduction to Operations Research’,

McGraw-Hill Education, 2021, 463. EXPLORING DECISION ADVANT AGES | 213 solution and so on. In each of the three iterations, the model considers all variables and constraints involved. Furthermore, Version 3 involves distance as a factor, by introducing two additional constraints (see constraints (5) and (6)). The first of the two defines that effectors can only be used if targets are within range in Stage One, given that the effectors are not aerial drones or aircrafts. This means that although an engagement option may be given in the compatibility table, the option is not possible if the target is out of range of Stage One and the effector is not an aerial drone or aircraft. The second distance constraints refers to Stage Two, expressing that effectors can only be used if target is within range in Stage Two and the effector is not an aerial drone or aircraft. The reasoning done by the model is that if an engagement is multi-stage, and if the target is out of range for an effector

in Stage Two and the effector is not an aerial drone or aircraft, then that engagement option is not possible. Version 4 introduces an additional layer of decision policy by introducing a differentiation between the targets. Although all targets handled by the OM are identified as being on the joint prioritized target list (JPTL), there could be rationales to handle them in a specific order. This scenario assumes that there are considerations given by a Commander within the targeting guidance asserting that there is an internal significance in-between the prioritized targets. Some of them are to be engaged before others. Version 4 considers all time-sensitive targets (TST) that emerge as priority 1, and high-value targets (HVT) as priority 2, and all component critical targets (CCT) are defined as priority 3. Version 4 thereby has five criteria in total. All versions were subject to the solution process which involved analysis of the results, refinements, and iterations. The solver

explored the feasible space (��) to efficiently search for optimal solutions (sets of attack option �� ) for each of the different scenarios. The progression from Version 1 up to 4 is the result of continuous analyses and refinements of the first model, where supplementary constraints, variables and changes to the problem were added to meet practical requirements and better reflect the real problem. Version 4 is considered the complete model (Model 2) to which the results, evaluations and conclusions are made against. 214 | EXPLORING DECISION ADVANTAGES Results The solution is multi-criteria optimal 100 with five (5) criteria. The solution time is done within seconds, and although internal metrics within CPLEX Solver states parts of seconds, it essentially takes a few seconds to read and comprehend the results as a human. An example of the complete scripting log from one example including the multi-objective solve log is provided in Appendix A. This example, or scenario,

contained 16 prioritized targets that required immediate recommendations on how to address. The following targets (Table 4) were included and, in this case, given an alternative priority based on their individual significance: Table 4: Targets that were used as inputs to the model (irrespective of version). Target 1 HVT C4I HQ Brigade Command Post Priority 1 Target 3 HVT Medium-Range SAM-system Priority 1 Target 5 TST Long-Range SAM-system Priority 1 Target 2 Target 4 Target 6 Target 7 Target 8 Target 9 Target 10 Target 11 Target 12 Target 13 Target 14 Target 15 Target 16 CCT CCT CCT HVT CCT HVT CCT TST HVT CCT HVT HVT CCT C4I HQ Battalion Command Post Short-Range SAM-system Airstrip Medium-Range Artillery Unit Short-Range Artillery Unit Attack Helo Site Tank Unit Priority 3 Priority 3 Priority 3 Priority 2 Priority 3 Priority 2 Priority 3 Long-Range SS Missile-system Priority 1 Medium Logistics depot Priority 3 Large Logistics depot Large Bridge Medium

Bridge Small Bridge Priority 2 Priority 1 Priority 2 Priority 3 100 A term used in here to define an outcome or result from a simulation. It refers to that the Solver could perform all the intended instructions given. In this experiment Model 2 is configured to first collect the data sets and then solve the problem according to the objective and parameters and given the constraints that conditions the solution. EXPLORING DECISION ADVANT AGES | 215 Out of the 16 targets, 5 were considered first priority, 4 as second priority targets, and 7 as third priority. This is visualized by the solver as: Solution (multi-objective optimal) with objective 5 Total number of Prio 1 targets killed is 5 Total number of Prio 2 targets killed is 4 Total number of targets killed is 16 Total cost in USD is 7655000 Model 2 first presents that it could optimally solve the multi-criteria problem, whilst considering all 5 criteria. After this it defines the division between the different targets in a

prioritized order. It also sums up the total cost, which in this example (scenario) came to a total cost of 7, 655,000 USD. In the following, the solution given by Model 2 includes the specific matching of the most appropriate effector and weapons type to each individual target. It also presents the specific attack option that it recommends, given all of the business rules, and targeting policy. This means that based on the data input the OM has considered the 16 different targets and respective features, the 13 different effectors with their respective features including the quantity of weapons available for each effector. In addition, the OM has used the compatibility data set of >330 options, and computed so that all 5 criteria were met, subject to 8 constraints that it had to consider before presenting its selected and recommended attack options: Target T1 is assigned to effector E5 with attack option A033 Target T2 is assigned to effector E5 with attack option A055 Target T3 is

assigned to effector E5 with attack option A075 Target T4 is assigned to effector E5 with attack option A085 Target T5 is assigned to effector E9 with attack option A090 Target T6 is assigned to effector E6 with attack option A093 Target T7 is assigned to effector E10 with attack option A107 Target T8 is assigned to effector E10 with attack option A132 Target T9 is assigned to effector E6 with attack option A153 Target T10 is assigned to effector E10 with attack option A183 Target T11 is assigned to effector E12 with attack option A213 Target T12 is assigned to effector E7 with attack option A231 Target T13 is assigned to effector E12 with attack option A264 Target T14 is assigned to effector E6 with attack option A294 Target T15 is assigned to effector E6 with attack option A318 Target T16 is assigned to effector E6 with attack option A330 Model 2 also summarizes the distribution of weapons used, since this was explicitly stated by the objective function (5) in the instructions. In

its solution, the solver describes the distribution between the available effectors, which enables a quick orientation of how the model chose to allocate its resources. Note that this can be 216 | EXPLORING DECISION ADVANTAGES manipulated by setting the quantity of weapons at a specific effector’s disposal to zero in the effector data set. It could also be used to reduce the equivalent number after each scenario, given that no replacements were made: Total units of weapons E1 used is 0 Total units of weapons E2 used is 0 Total units of weapons E3 used is 0 Total units of weapons E4 used is 0 Total units of weapons E5 used is 30 Total units of weapons E6 used is 21 Total units of weapons E7 used is 4 Total units of weapons E8 used is 0 Total units of weapons E9 used is 6 Total units of weapons E10 used is 12 Total units of weapons E11 used is 0 Total units of weapons E12 used is 7 Total units of weapons E13 used is 0 Evaluation The evaluation of Model 2 is multi-faceted. While

CPLEX Optimizer is renowned for its exact methods in solving Mixed-Integer Programming (MIP) problems (Hillier and Lieberman 2021, 463), having additional layers and sources within an evaluation reinforce the results. As seen in Figure 6, the evaluation and the integrated validations and verifications are used throughout the process concerned with evaluating the problem, the optimization model and the solution leading up to the results obtained from it. The evaluation serves multiple purposes, including ensuring the correctness of the model, the accuracy of the data, the appropriateness of the solver settings, and the feasibility and optimality of the solution. It is important to examine how changes in the data affect the model solution. Williams (2013, 238–40) refers to this as sensitivity analysis and a way to see if a model behaves in a stable fashion. He argues that the only really satisfactory method of performing a sensitivity analysis in integer programming problems involves

“solving the model again with changed coefficients and comparing optimal solutions” (Williams 2013, 239). It is a standard method for verifying simulation models The term ‘sensitivity analysis’ is, according to social scientists Chattoe, Saam and Möhring (2000), generally used to “describe a family of methods for altering the input values of the model in various ways and such analyses are included in the validation step of almost all technical simulations” (2000, 243) 101. Since CPLEX separates the model from the data (Watson and Cacioppi 2014, 9), performing a sensitivity analysis is essentially straightforward, either by switching between different data sets, or modifying some data in an existing data set without changing the model. Model 2 101 They also argue that there is a lack of methodological literature on sensitivity analysis in the social sciences. EXPLORING DECISION ADVANT AGES | 217 was found stable after completing the sensitivity analysis. Additionally,

building and using four models of altered complexity provided further insights to the robustness of the solution and the behavior under different circumstances (see the section Solution in this chapter). The feasibility of the solution was conducted by ensuring that all constraints, variables and objective functions were satisfied by the solution using CPLEX diagnostics and log tools (IBM Corporation 2013) and endorsements from IBM technical expert. The values in each data set were manually inspected to confirm that they made sense within the context of the problem. The performance index for evaluating the assignments was determined by four different criteria: the ability to propose feasible attack options of all targets to a decision-maker, the ability to handle prioritizations, the ability to perform re-attacks if the threshold for probability of kill (Pkill) is less than 0.9, and lastly to minimize the cost of the overall target engagement. These performance indexes were

operationalized through the objective functions. As is described within the results and visualized in Appendix A, all four criteria, or metrics of performance, were met by Model 2. It produces feasible attack options, including re-attacks, and can manage differentiation in prioritizations, as well as minimizing the total costs. Additionally, as seen in the multi-criteria functions, a fifth objective was included in Model 2’s objective function. This objective was deliberately added to try and ensure that the model strived to use as many different effectors as possible, making the solutions even more heterogenous. Outside of these criteria, the model was also evaluated using sensitivity analysis, CPLEX Optimizer’s internal validation protocols, and in-person verifications of all solutions. The capacity of the model to be utilized for cost-benefit purposes, capability development analyses and decision policies analyses was evaluated from comparisons between the different scenario

(or dynamic mission) results. These are conferred in the Section Discussion. No comparison was made from using other Solvers (software programs) to solve the same optimization problem (Hillier and Lieberman 2021). This decision was motivated by the experience gained in this experiment. The amount of time and competence required to perform a similar experiment using another solver was not appropriate given the lack of significant added value of doing it. The multi-faceted evaluation method ensured that the solution was not only mathematically optimal but more importantly, show potential to be a precursor in real-world applications. 102 102 The model was demonstrated for the Swedish Armed Forces as the main part of a proof-of-concept for a decision support system 9 October 2024. This was a continuation of the experiment in which the author and IBM developed a more user-friendly version and additional features. See Chapter 7, Extensions 218 | EXPLORING DECISION ADVANTAGES

Consideration of reliability and validity The most important factors when considering the reliability and validity of the experiment are arguably: (i) the interaction with the SMEs to understand the real problem for conceptual modelling prior to the model building; (ii) the data sets used as input for the model; and (iii) the reliability of the recommendations (output of the model). The factors strengthen the validity and reliability of the experiment, the model, and the results. External validations External validations from three directions were established. Despite the fact that the model is an idealized representation of the real world, validations of its performance from external sources are foreseen to be not only relevant but also enhancing the validity of the optimization problem and its solution as such. Given these incentives, three external resources have been engaged to complement the authors own evaluation: First, the CPLEX Optimizer includes features that test, debug,

and inform of any errors or in the case of finding an optimized solution, presents that the problem is solved without confliction (IBM Corporation 2013, 92,96,121,243); second, IBM provided a technical expert to review the results of the experiment including the CPLEX software performance; third, the use of subject matter experts in not only the conceptual and preprocessing phase but also from the interaction with end-users from the Swedish Armed Forces (subject-matter experts at a tactical and operational level of warfare). Besides these three resources, the experiment has been presented and discussed at the academic optimization conference ‘EURO24’. There are benefits of using CPLEX in relations to reliability and validity of the results. Like any software tool, IBM ILOG CPLEX Optimizer has its pros and cons when evaluated. However, the program is widely recognized for its integral capabilities and its efficiency in solving Mixed-Integer Programming (MIP) problems making it one

of the preferred choices for researchers (Hillier and Lieberman 2021; Lundgren, Rönnqvist, and Värbrand 2010). Moreover, the advantages of using CPLEX are also found in its speed, robustness, adaptability, and reliability. CPLEX optimization engine is powerful and can efficiently solve large-scale MIP problems fast. The use of a lexicographic function made it possible to create an internal priority feature where the solver worked its way through each of the objective functions in the order it was directed from the instructions given in the model code. No solver can solve multiple objectives at the same time. An objective needs to be a single minimum or maximum term in every mathematical model. Conversely, it is not completely straightforward to combine multiple criteria into a single set (as in this experiment), as criteria are not always compatible. For example, in this experiment combining EXPLORING DECISION ADVANT AGES | 219 criteria (1)(2)(3) referring to ‘maximizing kill

probability’ with criteria (4) ‘minimizing cost’, and (5) ‘maximizing the number of effectors utilized’ have different dimensions and scales. Lexicographic function is here a practical way to address multiple criteria with different scales and priorities. Each run finds the ‘global optima’ for the given objective. The lexicographic function therefore ensures that there are no ‘local optima’ within the set of multi criteria problem setting, making it invaluable to this research, since one of the key features is to enable targeting decision policies to be integral in this decision support application. Furthermore, the solver engine included various algorithms to support the solution, shown in Appendix A, where it presents its own cross-checking for potential issues, such as the need for relaxation. Since CPLEX separates the model from the data it is a simple task to change the data slightly and then run the model again to see how it reacted. Finally, it is dependable -

CPLEX has been rigorously tested and validated for years across a wide range of problems (Hillier and Lieberman 2021). In addition, CPLEX is accessible and free of charge to all researchers applying for access via their universities. On the other hand, the proprietary nature of CPLEX's algorithms could be a limitation if a researcher would need to modify the underlying algorithms. Limitations The results were never cross validated by using alternative solvers for the same problem. This obviously would have enhanced the reliability of the results On the other hand, the solver version used for the experimentation is defined. As is the mathematical formulations, allowing other researcher to reproduce the results accurately. Notwithstanding the absence of cross validations with alternative solvers, more than ten sensitivity analyses were conducted to understand the impact of changes to the model input and parameters. This is thought to have enhanced the validity by assessing the

robustness of the solution against changes in the model. By beginning with a simplified version, this step-by-step validation helped in isolating and understanding the contributions of different components of the model. Additionally, the experiment was externally validated by the use of two external sources. First, an optimization expert from IBM performed an independent verification of both the construction of the model and the subsequent results. Secondly, military subject matter experts were engaged in validating the data sets used, including the compatibility matching between each target and effector to ensure that any combination would be not only feasible, but a recommended solution. The outreach to these two external sources is recognized as a real-world validation to the experiment as a whole. 220 | EXPLORING DECISION ADVANT AGES Discussion The last section of the chapter discusses the experiment, the results, and the implications of Model 2, including an analysis using

the operationalized theoretical tool (Table 2 in Chapter 3) and a summarization of the chapter. The lexicographic function The types of problem presented can be perceived as a multi-criteria problem where there are several different criteria to fulfil, and where there are some kinds of priority ranking in-between these. Such problems are classified as ‘lexicographic multi-objective problems’ (Cococcioni, Pappalardo, and Sergeyev 2018, 298), as previously described. The order in which the five arguments (or criteria), that comprised the multi-criteria function, where prioritized followed the logic of the US and NATO targeting doctrine. The chosen order is one of many ways to order the priority of the arguments (or criteria), all of which are to be satisfied using a lexicographic approach. It is possible to argue for a different order or priority ranking, for instance, to put cost minimization before minimizing the number of weapons. However, two arguments are provided here, to

provide some rationale for prioritizing the minimization of weapon usage over cost minimization in this framework: (1) Operational efficiency: Minimizing the number of weapons used is essential for conserving critical resources for future targeting engagements. By prioritizing weapon efficiency, the model ensures that the joint force retains a sufficient supply for various operational demands. Furthermore, it contributes to mission sustainability and long-term operational effectiveness, which arguably has a higher priority than immediate cost savings. (2) Targeting effectiveness: Focusing on using fewer weapons can lead to more precise targeting and efficient strike strategies, which may (by pure probability calculation) reduce the risk of collateral damage. In contrast, placing minimizing cost as a higher priority than minimization of weapon usage could lead to using the minimum viable resources, potentially compromising mission effectiveness or increasing operational risk.

Prioritizing weapon efficiency arguably aligns with achieving a commander’s objectives more reliably than focusing on cost. These arguments are also the rationale for lexicographically ordering weapon minimization above cost considerations, emphasizing mission-critical needs over financial optimization in Model 2. Implications and value of Model 2 The study aimed to show how optimization models (OM) can be applied to augment human decision-making in a dynamic targeting process. The results reveal that the OM solves the problem and finds a multi-criteria optimal solution within seconds, given the conditions defined in the optimization model. It dramatically compresses EXPLORING DECISION ADVANT AGES | 221 the OODA loop and the time it takes to complete a dynamic targeting cycle. This has great significance for the application of warfare. Model 2 is valuable for several purposes. First, incorporates multi-criteria functions that permit prioritization in-between the targets given

their contextual significance. It also accounts for minimizing the total number of weapons used, thereby trying to use the most appropriate number of weapons given the preferred effect, which in these scenarios was defined as destroy. Second, as all target engagements were related to creating the specific effect of destroying the function of each target, a precise metric for probability of kill (Pkill) was introduced in the model. The probability calculations were linearized and made dependent on any earlier stage’s battle damage assessment and re-attack recommendation. Based on subject matter experts’ declarations on the appropriate number of weapons for each feasible weapon that was required to meet a 90% probability of kill, the model was given three optional protection attributes to each target. Third, all targets were predefined in each scenario as to their protection level. Either they were declared as being vulnerable, neutral, or protected. This parameter was considered an

additional source of target intelligence, in managing the uncertainties surrounding their actual preparations and readiness, for example, the use of camouflage, decoys, and other protective measures. Moreover, and in reference to the effectors and their respective weapons, the Pkill value of protected targets also accounts for each weapon’s potential malfunctions or miscalculations in the desired point of impact for each target. The probability of kill function (Pkill) creates a logical path towards a multistage target engagement, in which the second attack is related to a re-attack on the same target to ensure the desired effect is met. The model can be perceived as a precursor for future developments of decision support systems (DSS) in targeting practices. Model 2 was acting based on the following logics and input data. The effector data set is perceived as data communications between the model and the variety of joint effectors available and interconnected in a joint force

network of targeting services – constantly updated to allow for distributed targeting from a dispersed but interlinked kill web. Similarly, the target data set is perceived as being fed into the model via the intelligence function using the joint forces sensor network and correlating the emergent target with the joint force targeting database. Both of these ‘services’ are non-existent in the real-world, to the best of this project’s knowledge, and have therefore played a fictional part in the set-up. The third data set is the compatibility table, which is not a fictional input. This kind of data set can be produced to cover all options foreseen to exist between the two other data sets, and then integrated in an optimization model. 16 targets of significance that posed a variety of threat to the fictional joint force were dealt with using 13 different joint effectors. It led to >300 attack options in the compatibility data set. 16 prioritized targets at the same time are not

seen as a small number. On the contrary, the number is likely a bit exaggerated The number 222 | EXPLORING DECISION ADVANT AGES of joint effectors was less important than ensuring a rich enough variety of effectors. The OM could choose between short to long range artillery, ground-based missile systems, armed UAS, to fighters equipped with different munitions. Additionally, the sensitivity analyses (which comprised a step-by-step progression, adding complexities along the way, and comparisons from changing data and parameters in the respective data sets), added value to the overall reliability. Analysis In reference to Table 5 below (and in Chapter 3, Operationalization), the optimization model (OM) was augmenting military decision-making within a dynamic targeting setting. As such, it was engaged in 11 different, doctrinally derived subtasks when it solved its problem, which are marked by colors in Table 5, column 4 and 5. Neither of these can be seen as minor to another. Their

individual significance has two faces One is the complexity of the subtask itself; another is the way the subtask is represented mathematically. Table 5 The table provides an overview of the different tasks that Model 2 contributes to, as well as its relational engagement within the JTC and the OODA loop. OODA Observe Orient Decide JTC Phase 1 Cmr’s Intent, Objectives, and guidance Phase 2 Target development Phase 3 Capability analysis Phase 4 Dyn TGT Find # Subtasks/Functions 1 Search 2 3 4 5 Fix 6 7 8 9 10 Track Target 11 12 13 14 Detect Identify Determine TGT characterization Classify Confirm TGT validation Significance PID TGT mensuration Prioritization Monitor Estimate TGT window of vulnerability Risk estimation Desired effects Intel support activities Indicators & warnings JIPOE MoP/ MoE Target System Analysis Target Material Production Basic target development Estimations Intermediate target development ISR management Functional characterization

Damage estimation # 1 2 3 4 5 6 7 8 9 10 11 12 EXPLORING DECISION ADVANT AGES | 223 Cmr’s decision, Force assignments Phase 5 Mission planning 15 Time management 16 RoE/TGT Policies 17 Collateral damage Estimation (CDE) 18 Threat analysis 20 Sensor allocation 15 26 Recommendations Requirements incl Combat assessment Deconfliction/Coord/ Synchronization Select option Issue order CID Field CDE 28 Measurements & Documentation Collection tasking 16 29 Follow-on actions 30 Re-attack Battle damage assessment (BDA 1) 17 19 21 22 Force execution Engage Exploit Act Assessment Assess 224 | EXPLORING DECISION ADVANT AGES Functional analysis including Advanced target development 23 24 25 27 Options Target engagement 31 Combat Assessment 33 Re-attack recommendations 32 MEA Threat analysis 13 Intel requirements 14 BDA 2 Re-attack recommendations BDA 3 18 19 20 Interrelation to the OODA loop The experiment interrelates to all of

Boyd’s four parts as visualized in Table 5. Acting as the theoretical foundation to which the results can be related and explained, the traditional kill chain portrayed in F2T2E2A, and further derived to 33 subtasks or functions, is to a significant extent its own OODA loop. By applying an AI-application that engages in 11 subtasks, it can be argued that Model 2 empirically and theoretically resides in OODA. The first O (Observe) is not related to the derivations made in the subtasks of the F2T2E2A. Nonetheless, the data sets of the targets and effectors that the model considers are both seen as elements of Observe. So is Boyd’s feedback loop from Act to Observe. It correlated to the initial observations on the desired effect after the first stage of attack. If 09 in Pkill was not achieved, then the model automatically decided to make a re-attack recommendation. Comprehension and speed within a system corresponds to the scope of the OODA loop. The faster a system can retrieve data

relevant to a problem, make sense of it, select feasible options, decide, and act, the better (all other things being equal). This tallies the main framing of the contemporary targeting concept, though this experiment implies how this optimization model could reside within a larger targeting web (see Chapter 7). The results suggest that an optimization model can augment decision-making within a dynamic targeting setting by selecting optimal solution as recommendations from data in just a few seconds. AI augmenting the joint targeting cycle The experiment is also a manifestation of how an intelligent agent acts in support of the joint targeting cycle. As illustrated in Table 5, the agent (Model 2) is more or less engaged in all six phases of the JTC when solving its problem. In phase 1 it considers and implements the targeting guidance as either variables, objective functions, or constraints, to include how to prioritize in-between targets of various significance, minimizing the cost

and maximizing the utility of different effectors. These are all viewed as components of the decision policy, integral to the attack options and the recommendations given by the OM to the fictional human decision-maker. By introducing an approach to incorporate a commander’s intent, objectives and guidance in the optimization algorithms, a decision policy can be mathematically formulated and applied in decision support tools. Additionally, Model 2 will adhere to the policy (or business rules) every single time, until they are changed, and subsequently verified as effectively reaching an optimal solution. If Model 2 encountered any inherent problems within the business rules, for example that it could not meet one of the criteria, it would then inform the decision-maker of this in its output (see the example of a solution on page 127). EXPLORING DECISION ADVANT AGES | 225 Its main contribution to the joint targeting cycle (JTC) is its augmentation of the available options and its

recommended decision in phases 4 and 5, which could be valuable also in a less time-compressed situation. If decision-makers can retrieve feasible attack options at any given time that they request it, then that is believed to be valuable information prior to any targeting decision, force assignment, or mission planning on both a tactical and operational level of command. A specific notion needs to be attended when referencing JTC’s phases 4 and 5, or for that matter between Target and Engage within F2T2E2A (see Chapter 3). If Model 2 would be interconnected with all effectors in a joint force network, it would know their status and availability, then the findings of this experiment indicate that all of Model 2’s recommendations automatically would be synchronized in time and availability. Since the experiment assumes this within its simplified problem by directing the Model to read the data from the effector data set, no coordination would be needed either, since the status of a

specific effector would read ready to serve a prioritized target. A complementing feedback loop by an instruction in Model 2 coding would enable it to instantly update the weapons status within the network. Deconfliction between own forces can also be managed by integrating the inputs from the common tactical pictures to a common operational picture (COP) that are in use already today to serve a commander’s situational awareness. The speed of targeting recommendations The speed in which the model can produce targeting recommendations that follows predefined rules of engagement is significant. The model solved each problem in seconds, and sometimes within parts of a second. As the experiment used an exact method (CPLEX Solver) each optimal solution is precise, efficient, and therefore consistent. In comparison to the current practices (baseline condition), where several targeting cells would have to coordinate themselves to solve the same problem, this is assumed to take a lot longer

time. Considering that a contemporary joint task force would have at least a targeting cell at the operational level and one targeting cell at each component command, this would be a coordination between four (or more) cells working to figure out how to match the targets most optimally with what effectors and weapons. They would also have to consider which target to prioritize first, second and so forth, as well as safeguarding adherence to the targeting policy. In this case that is to consider five criteria, subject to the eight constraints for each target engagement option they would recommend. This is of course possible, though not in seconds. Moreover, given that there were 16 targets that required attention for rapid decision-making, the selected option would likely not adhere to the defined targeting policy in each attack option. 226 | EXPLORING DECISION ADVANT AGES Performance as measurement The solution time is a clear measurement of performance. The performance index for

evaluating the assignments was determined by four different criteria: the ability to propose feasible attack options of all targets to a decision-maker, the ability to handle prioritizations, the ability to perform re-attacks, and to minimize the cost of the overall target engagement. All four criteria were met by the model The dynamic weapon to target assignment (DWTA) problem involved multi-criteria criteria, technical and operational constraints. The mission (scenario) was considered complete if all targets were engaged and the desired effect (destroy) was accomplished for each target, subject to the multi-criteria functions and the constraints given to the model. Additional advantages There are some additional advantages that increase the utility of the model. While the study provides evidence for the effectiveness of optimization algorithms to predetermined WTA-problems within a dynamic targeting environment, it also addresses other shared challenges that could be of use in a

larger decision-making framework. First, as the model incorporated cost and benefit computations, this can be used to efficiently calculate the cumulative expenses of, in this case, targeting activities. When introducing the objective function of minimizing the total cost of a solution, the OM was able to meet this objective and reduce the total cost by millions of dollars. The cumulative effects of such an approach would make an impact eventually on the economy. The experiment used a budgetary regulator Each scenario was given a specified budget restraint. By regulating the budget, changes in the model’s choice of attack options were exposed. It suggests that it is possible to integrate the principle of economy of force into a WTA optimization algorithm. Second, the lexicographical approach makes the model a suitable framework for an armed force’s modelling and simulation activities for capability and investment planning. By adding new effectors, or subtracting others, one can

arrange a scenario that simulates if a specific capability (effector or weapons type) is ‘preferred’ or not compared to others, given a set of conditions in a controlled environment. Analytics from such simulation activities could also be of value for wargaming, and for improving the understanding of how different criteria and constraints can co-exist or be in conflict with each other. EXPLORING DECISION ADVANT AGES | 227 Conclusions The experiment covered the exploration of an AI-model built to optimize the best possible decision given its input data. The experiment evidence for the benefits of machine reasoning using optimization algorithms to improve decision-making in targeting. It claims to have used a representative dataset build on corollaries from discussion with subject matter experts in weaponeering and intelligence within the Swedish Armed Forces. The fundamental AI technique applied to the problem was optimization. The problem was managed using integer linear

programming. The use of external military subject matter experts to suggest and validate the attack options as well as the weapons and target data, was an attempt to make the experiment authentic and accurate. The AI model that was built used an exact method, which makes its solutions precise (given that the problem is solvable) rather than less accurate (if a heuristic method had been chosen). The model (OM or Model 2) was able to solve all problems, given all of the business rules and targeting policy, specifically recommending the most appropriate effector and weapons type to each individual target. In each simulation, the OM considered 16 different targets, 13 different effectors with a total of >330 attack options. It computed (machine reasoning) so that all criteria and constraints were met within seconds before presenting its recommendation. The main advantage of the model is that it is rule-based. It can (in theory) integrate commander’s intent, objectives, and guidance

into algorithms that represent the commander’s decision policy to targeting. Dynamic targeting necessitates speed and precision. The evaluation indicates that the OM can provide optimal solutions that are precise, efficient, and consistent. The performance index was determined by four different criteria: the ability to propose feasible attack options of all targets to a decision-maker, the ability to handle prioritizations, the ability to perform reattacks, and to minimize the cost of the overall target engagement. All four criteria were met by the model. Its main contribution to the joint targeting cycle (JTC) is its augmentation of the available options and its recommendation of the best possible decision. Rapid comprehension within a system also corresponds to the scope of the OODA loop. The faster a system can retrieve data relevant to a problem, make sense of it, select feasible options, decide, and act, the better (all other things being equal). Although the intelligent agent

in this experiment is limited in its ability to comprehend any context beyond the data set provided, more data could be consumed by it. However, context awareness is an area where future work and research is required to further explore how optimization models can be made more prone to context, changes, and consequences in a targeting environment. 228 | EXPLORING DECISION ADVANT AGES Consolidations of such features would generate AI systems becoming synthesized targeting advisors offering a congregation of targeting services to targeting directorates. This is further discussed in the next, closing chapter, where conclusions from the complete work are covered. EXPLORING DECISION ADVANT AGES | 229 230 | EXPLORING DECISION ADVANT AGES EXPLORING DECISION ADVANT AGES | 231 Chapter 7 – Conclusions Introduction The purpose of the project is to investigate the integration of artificial intelligence (AI) into military targeting processes to augment critical decision-making,

thereby enhancing speed, precision, and consistency. A theoretical framework is constructed to provide a common reference point for the two empirical investigations. The project uses a multi-methods approach centered around modelling and simulation to elicit the benefits of AI to create decision advantages under specified sets of conditions. It involves two experiments each configured to expose contemporary problems within joint targeting and to propose new ways to solve them using AI. The AI-models either imitates and substitutes human tasks or address the problem in new ways. The results present novel perspectives on applying AI to joint targeting The chapter begins by addressing each objective stated in the thesis and a summary of the key findings or a reference point to where each objective is discussed. It focuses on the core results from the two experiments. By doing so, it lays the foundation for discussing the research contributions and its implications. The chapter continues

by presenting the contributions of this research to the existing body of knowledge. It highlights the novelty of the work, and the advancements made through this research. It explores the implications of the findings addressing their potential influence on method, theory, practice, and strategy (policy). It separates contributions from implications to provide clarity between what has been achieved from what this may influence. These are further divided into methodological, theoretical, practical, and strategic contributions and implications. The fact that the design and innovative use of methods produces the results motivates the order of appearance. The chapter discusses the limitations of the study followed by directions for future research and suggests areas of unresolved issues and opportunities that have arisen from this project. It completes its undertakings by providing a final reflection on the broader significance of the project. 232 | EXPLORING DECISION ADVANT AGES

Summary of key findings Objective 1: To develop and apply two AI models that address specific problems in contemporary joint targeting to demonstrate practical applications of AI in this context. Model 1: Enhancing precision and efficiency in a joint force dynamic sensor allocation and target engagement with deep learning. The main advantage of the model is its augmentation of intelligence support activities when applied in the dynamic targeting process. The analysis shows that it is supportive of 9 out of 20 intelligence support activities. Most importantly, Model 1 enhances ISR management (#9), Sensor allocation (#15) and Collection tasking (#16) making the use of a military force’s own sensors more effective and efficient. The optimization of sensors also facilitates faster target engagement. The first performance criterion is the ability of a neural network to mimic and learn to emulate signal attenuation of a TER radar. By creating a synthetically generated training environment

using satellite images from SRTM-1 and some commercial programs for statistical calculations of RF signal propagation loss due to topography, the neural network could be trained and validated (Fig. 23 and Table 1-3) The second and third criteria are also met, since the U-Net improved its ability to emulate and infer along the axis of iterations (Fig. 23) and the accuracy of the model’s output could be verified (Table 1). Lastly, the fourth criterion of the performance index is the ability to reduce the time compared to a statistical program, which was verified (Table 2,3). In addition to these four criteria, the evaluation of Model 1’s performance is also validated by using an external group of human subject matter experts. This was done by giving five SMEs an individual task of suggesting the preferred site location of the same type of TER in the same areas that were used for simulations, and subject to the same conditions that the model was given. The use of an external group of

human SMEs made it possible to superimpose their respective suggested solutions onto the model’s solutions and to compare the results for evaluation purposes. Figures 30 and 31 illustrate two of the simulations where the SMEs suggestions were superimposed to Model 1’s predictions. The correspondence (Fig 30 and 31) is indicative for good predictions. Furthermore, the evaluation of Model 1’s quality of inference reveals an IoU-scoring of 0.77 at the end of the training The inference time (the task completion of ResNet34) was up to 40 times faster compared to one of the more frequently used calculation models (Splat!). After finetuning Model 1 (using K-means clustering) it could correctly infer the most likely positions of a medium-range surface-to-air missile systems (MSAMS) target engagement radars’ (TERs) sites, given the EXPLORING DECISION ADVANT AGES | 233 constraint defined in its instructions. The speed, accuracy and consistency in which it could provide its

recommendations of these high value targets (HVTs) enhances the precision and efficiency in a joint force dynamic sensor allocation and target engagement. Model 2: Optimizing a Joint Force’s Dynamic Targeting Decisions with machine reasoning. The main advantage of this model is its rapid multi-criteria recommendation of attack options applied in the dynamic targeting process. The analysis shows that it is engaged in 11 (out of 33) different and doctrinally derived subtasks of the F2T2E2A method. Most importantly, Model 2 optimizes the best possible options available given its input data. The speed at which it can produce targeting recommendations that follow predefined rules of engagement is significant. It solved each problem in seconds and sometimes within parts of a second. As the experiment used an exact method (CPLEX Solver) each optimal solution is precise, efficient, and therefore consistent. The results suggest that a multi-criteria rule-based optimization model can improve

efficiency in decision-making within a dynamic targeting setting by selecting optimal solutions as recommendations from data in just a few seconds. The performance index for evaluating Model 2 solutions was determined by four different criteria: the ability to propose feasible attack options of all targets to a decision-maker, the ability to handle prioritizations, the ability to perform re-attacks and to minimize the cost of the overall target engagement. All four criteria were met by the model. The solution involved multi-criteria aspects, technical and operational constraints. The mission (scenario) was considered complete if all targets were engaged and the desired effect (destroy) was accomplished for each target, subject to the multicriteria aspects and the constraints given to the model. Objective 2: To analyze how AI augmentation can expand the decision space for military commanders to facilitate a shift from hierarchical and linear targeting structures to dynamic and

non-linear concepts. See section Practical implications and Strategic contribution and implications Objective 3: To assess the implications of AI integration in military decision-making processes, command and control arrangements and joint warfare concepts. See section Theoretical implications. Objective 4: To provide empirical evidence and practical insights for practitioners on how AI applications can augment joint targeting practices. See Objective 1 and section Practical contribution. 234 | EXPLORING DECISION ADVANT AGES Objective 5: To contribute to the academic discourse by exploring the transformative potential of AI in military contexts and recommending areas for future research that include ethical considerations and impacts on command authority. See section Theoretical contribution and implications and Directions for future research. Methodological contributions: Using a multi method approach to solve the problem The project used a multi method approach and is a

contribution to the developing literature in War Studies. It partially aligns with Biddle’s (2010) innovative use of multi-method analysis in War Studies but places a stronger emphasis on addressing the specific military problem that arise from applied AI research. By constructing and applying two models designed to augment the dynamic targeting method, the study showcases practical applications of AI in this context. The design of the project imposed an interdisciplinary methodological approach of mixing military practical knowledge with strategic theory, military doctrines, and technological knowledge. This strategy allowed a multifaceted, multi method analyses and subsequent synthesis that facilitated novel insights, innovative findings and externally recognized utility. Using experiments for this explorative research was motivated by the research problem and the scope of the project. It is located at the heart of joint warfare The problem concerned the integration of AI into the

joint targeting process and the associated decision-making. The way in which AI-applications could be empirically tested and evaluated required modelling and simulation, similar to the requirement laid out by Biddle in his research on military power (2010, 10–11). The simulation environment enabled a controlled environment in which a real problem could be simplified, modelled, simulated and evaluated. The guidance for this was influenced by operations research (OR) literature (Hillier and Lieberman 2021; Williams 2013; Stewart Robinson 2017; Lundgren, Rönnqvist, and Värbrand 2010). From a methodological perspective, it shares the same perspective on experiments as found in the descriptions on the matter within social science (Webster and Sell 2014; L. Cohen, Lawrence, and Morrison 2011). However, the descriptions on how to do modelling and simulation were richer in the OR literature and the methods literature in social science were more related to problems concerning humans. The

fact that new methods of using AI for experiments in social science are becoming accessible (Green and White 2023) indicates that the field is underdeveloped in this context. The experiments concern real-world problems but involve interventions where AI aims to solve targeting-related problems. It is neither feasible, nor legally or ethically justifiable to run these experiments as fielded experiments. Consequently, a EXPLORING DECISION ADVANT AGES | 235 simulated environment was created to which performance metrics were developed to measure the effects of the interventions. The criteria of performance were concise and could be supported using standardized measurements such as IoU (for Model 1) or software such as CPLEX control tools (for Model 2). As it turned out, these performance criteria became useful tools in evaluating the performance and utility of the two models. Both experiments involve external experts. This supports the construction of the real problem into a simplified

problem and the data used. Their engagements also increase the validity of the findings. Most importantly, the attempt to be as detailed and transparent as possible in how the two were constructed and conducted enables both academic scrutiny and replicability. The controlled simulations permit both comparison across different settings and reproduction for other researchers. As a methodological reflection, the overall approach and the interventions in which “non-traditional tools” (in this case AI) was applied shares resemblance with Boyd’s OODA loop and the way “we derive knowledge from our environment” (Osinga 2007, 242). Using innovative techniques and methods in the two experiments The predictive analysis capability of Model 1 was achieved by using neural networks for deep learning. The intelligent agent learnt from being supervised to make predictions. The summary of key findings for Model 1 indicate the utility of the model for sensor optimization. It enhances sensor

allocation and intelligence collection The strategy to use subject-matter experts added valuable knowledge to the modelling phase and reliability and validity in the evaluation of the results from the simulations. The survey conducted facilitated not only external verifications of Model 1’s predictions but also made it possible to fine-tune the model. The postsurvey analysis also enabled a subsequent fine-tuning of the model in which the model was given an additional task to optimize the nominated target candidates it provides in Step 3. In this add-on the model considers the balance (K-means clustering) of the cluster size relative to the minimum distance between each cluster and each reference point (being the human experts’ input). As a final remark, it is worth noting that the method applied (semantic segmentation of high-resolution satellite images in which a neural network learns how topography affects RF signal attenuation) has a patent application to it. The second

experiment uses optimization algorithms to compute the best possible options given its data input. This rule-based method involves mathematical decision integrations from implementing multi-criteria aspects within its objective as well as constraints. The choice of using a lexicographic function within the model to manage 236 | EXPLORING DECISION ADVANTAGES this was successful and showed potential. The problems presented to Model 2 in each simulation were multi-criteria problems where there are several different criteria to fulfill and where priority ranking exists in-between. These types of problems are classified as ‘lexicographic multi-objective problems’ (Cococcioni, Pappalardo, and Sergeyev 2018, 298). The order in which the five criteria (or arguments) were prioritized followed the logic of the US and NATO targeting doctrine. The chosen order is one of many ways to order the priority of the arguments (or criteria) and all of which are satisfied using a lexicographic

approach. It is possible to argue for a different order or priority ranking; however, the methodological contribution refers to the inference that can be made. It shows that decision policies can be mathematically transformed into multi-criteria aspects of targeting decisions. The evaluation of the method and techniques involved are seen under the section Key findings. Another important aspect is its contribution to orchestrate decisions under time and resource constraints. The model’s ability to consider numerous targets and effectors and to compose rule-based target engagement recommendations in just a few seconds was a vital part of the conceptualization and construction of the model. The method showcases a tangible example of multi-criteria rule-based optimization model. Both of the methods advance state of the art research by providing tangible examples of applied AI for military purposes and contribute to the exiting literature on military AI applications further discussed in

the subsequent sections. Methodological implications: Simulations forge science and practice into applied research The findings have implications for applied research within War Studies and the wider social sciences. Each experiment begins by defining a real-world problem that was simplified and conditioned to elevate the innovation. Both simulations demonstrate how applied AI can expand and reinforce human decision-making making the targeting process more efficient and rapid by three evident features. First, they both treat quantitatively large data input to produce qualitatively actionable information. Their tasks are challenging or even beyond human cognitive limits in terms of speed, precision and consistency. Second, the synthetically produced knowledge postulates decision advantages for a targeting enterprise if similar AI systems were to be implemented. Third, the benefits from having these AI systems compared to not having them are showcased in the key findings and when

reviewing their respective contributions. Simulations should arguably be used more frequently for research in War Studies and the social sciences. They offer at least three distinct advantages These include (1) Controlled experimentation because modelling and simulation allow researchers to test theories and scenarios in a controlled environment and enable the EXPLORING DECISION ADVANT AGES | 237 exploration of complex phenomena without the ethical or logistical constraints of real-world experimentation. (2) Predictive insights as they provide predictive capabilities by simulating outcomes of various strategies or policies thereby simplifying the understanding of potential implications and guiding decisionmaking in contexts that otherwise can be demanding to explore. (3) Interdisciplinary integration as modelling and simulation bridge qualitative and quantitative approaches and facilitate the integration of diverse data sources and methods to address multifaceted problems.

Theoretical contributions: Applying, extending and challenging existing theories The research problem required a scientific baseline for military decision-making. A theoretical framework was built based on Boyd’s ideas that comprises the OODA loop and the joint targeting cycle (JTC) theoretically defined in NATO and US doctrines. Both the OODA loop and the JTC logically stem from a human-centric approach. Human comprehension of a situation is based on interpretations of observations, which is the case also for the staff’s estimations, the commander’s assessment and the intelligent agent’s predictions or recommendations. They all share the same sequential logic of perceiving before acting. Correspondingly, the differences in sensory input (keyboard strokes, data files, or briefing) do not change the requirement of establishing an awareness of a situation prior to deciding and acting. This is an important aspect of decision-making in this thesis Boyd’s theory was applied for

decision-making at the joint level of warfare with an upper boundary towards associated concepts for command and control. However, this is not unique. A review of the US DoD strategy on command and control for alldomains operations (JADC2) articulates the three guiding functions of C2 as ‘sense,’ ‘make sense’ and ‘act’ (2022, 4). Nonetheless, the OODA loop formed the foundation for the JTC, the method of dynamic targeting and the intelligence support activities. Collectively, they could be operationalized and interlinked to support the empirical investigation. However, the observed reality or the starting point for this study was that military organizations were thought to need AI to overcome contemporary targeting challenges. An integration of AI could theoretically either solve an existing problem, improve existing process or even create changes to the targeting concepts. The experiments were built upon existing real problems but tried to avoid the current conceptions of

how these are solved. This approach opened for novelty and explorations of AI employment. Integrations of AI will differ depending on the specific context. Understanding the contemporary and future targeting environment is vital. The targeting process will continue to reside within an OODA loop and 238 | EXPLORING DECISION ADVANT AGES thereby represent a race measured against an adversary’s corresponding process. Understanding each context and having the means to act are essential. To outperform an adversary's OODA loop requires either slowing their cycle using conflicting information / deception or accelerating our own by enhancing decisionmaking through knowledge consolidation. AI offers significant potential to augment, replace or collaborate within the human OODA loop as demonstrated in the experiments and their extensions. As new knowledge emerges, the opportunities and use cases for AI in this context will continue to expand. What has been discussed here advances

existing theories by extending and challenging them. This is further exemplified in the next section where Boyd’s OODA loop is challenged by illustrating how AI augments each stage in the decision-making calculus. Theoretical implications: Reshaping the theoretical discussion and the current decision-making paradigm The implications of this project reshape the theoretical discussions concerned with evolving human-machine relationship and the current decision-making paradigm. The implications discussed are built on the results and the methodologies used in the study. Even if the project has no intention of developing Boyd’s theory as such the investigation still has implications for the theory in terms of extending and challenging the ideas behind it as well as other military decision-making frameworks, for example, the joint targeting process. However, this discussion showcases the OODA loop as it is the theory used in the study. Expanding on the four basic components of the OODA

loop from two simple perspectives of time (at time of origin and now), it is apparent that computers and subsequently AI are becoming more dominant in contemporary military employment of the OODA loop (See Table 6). EXPLORING DECISION ADVANT AGES | 239 Table 6 OODA Loop OODA Before (at time of its origin) Now Trend Observe Primarily done by human senses, apart from signal surveillance Most observations are done by ‘digital senses’ (ex: LLMs searching the Internet, computerized vision) Technology taking over is Orient Most information is data, making automation and computerized analytics increasingly supportive to this part of the loop. Towards awareness AI Decide Primarily done by human analysts that processed the information and made assessments, assumptions, deductions, and conclusions. Exclusively humans Predominately humans Act Execution by machines existed. Intelligent machines that act continue to grow in numbers and sophistication (e.g drones, fire

control in weapon systems) More augmentation by AI, and automation and autonomy through AI (e.g satellites) Towards AIenabled sensorsand weapon systems Beginning with Observe, the contemporary application of Boyd’s theoretical framework indicates a strong acceptance for military employment of non-human sensors to observe (and collect) the data and information that is perceived as necessary to make sense of a situation. The use of the Internet and web-services, in combination with the use of military sensors and computer technologies represent a present-day approach to observe. The empirical findings underscore this and show how such an approach can be even further exploited. In particular the results from the first experiment, where Model 1 can be perceived as a forerunner for 240 | EXPLORING DECISION ADVANT AGES synthetically produced estimates of an adversary. Sensors are an important means of detecting and acquiring external information about targets. Physical targets have

detectable physical attributes, including shape, speed, vibration, acoustic noise, or signals emitted and reflected. Non-physical targets such as ‘virtual targets’ (NSO 2021, Edition B:24), for instance, a web-site have no physical attributes and though they can be observed by humans, this can only be done through the use of computers and software. Nonetheless, both physical and non-physical targets require detection prior to target engagement. Military sensors can locate a target, but they also assist in tracking a target and can guide weapons used to its desired point of impact. The sheer amount of accessible data necessitates computational power and automated processes within Orient. The capacity for inference and diverse analytics from acquired data is vast and continuously evolving. Human comprehension involves understanding what is happening, determining appropriate actions and deciding how, when, and with what resources. Today, this requires computer assistance. Software is

essential to filter irrelevant information and prioritize critical insights. While access to valuable information is no longer the primary challenge, the ability to manage a variety of domains, distill data and distribute actionable intelligence across a joint force in a timely manner remains a pressing issue. As revealed in a historical study of the Vietnam War, the size of US information flow impacted its decision-making (van Creveld 1989, 246–47). As a deduction from this, the management of the amount of accessible data today using human analysts exceeds any reasonable consideration. Importantly, the results of the second experiment disclose that to orient (comprehend) prior to making informed decisions under dynamic conditions necessitates AI augmentation. However, if the first two parts of the OODA loop (observe and orient) are increasingly accommodated by machines with AI, then human decision-makers may find themselves deprived from having the foundations for making informed

decisions. Boyd himself perceived orient as the most important part of the loop (Osinga 2007, 192). This will condition the two subsequent stages in the loop. How will commanders and other decision-makers judge the data provided by intelligent though synthetic actors prior to decision? Human judgement is still vital. However, the trajectory of context-aware AI may have its next stop at cognitive AI. An artificial awareness beyond the context-awareness of today, could co-create non-human intelligent agents with an ability to assume professional judgement. Professional judgement is an important ingredient when applying practical knowledge and conducting warfare as it guides the “decision-maker when faced with novel situations, including in handling the unknown” (Bovet Emanuel 2024, 253). However, and importantly, professional judgement does not exclude nonhuman agency It is important to consider the amount of time it takes for a military commander to gain experience and practical

knowledge as the foundation for their ability to judge. EXPLORING DECISION ADVANT AGES | 241 Contrast that with a cognizant AI-agent trained for AI-based decision-making at machine speed in a defense simulation environment (or metaverse) in a way that builds its professional judgement over time. The agent could synthetically simulate millions of missions having a command position, thereby gaining a magnitude of more experience than a human could ever attain in a much more compressed timeframe. Additionally, if such AI-agents were trained against simulated adversarial AI-agents, then new tactics and strategic approaches not thought of could evolve. However, this would require fundamental change in a nations “strategic culture and organizational set up” as suggested in a study on Germany’s defense AI approach (Borchert, Schutz, and Verbovszky 2024, 212). Nonetheless, the primacy of human decision-making is challenged. Furthermore, some warfighting concepts even prerequisites AI

in a decision-making position, for example, DARPA’s mosaic warfare (2018b) and modern ‘software-defined defence’ (Soare, Singh, and Nouwens 2023, 8). The last stage of the loop is Act. Machines can technically execute decisions as a direct extension of an explicit order. However, during the last few years more implicit and independent actions are seen in a study of the development of autonomous air defense system (AADS) that automatically detect aerial threats and target them without human intervention (Khan et al. 2021, 1) Defense industries, including Rheinmetall, Rafael, Raytheon, and Anduril all develop AADS and these are becoming more sophisticated as Khan et al. suggests (2021) Contemporary research is seeking to advance combinatory approaches where the AI system consists of combinations of AI-models, each with its specific role, as Du et al. suggests with their CA-MAS (2024). This could enable giving complete missions to an AI system, that subsequently derives logical

subtasks given the intent of the mission, assigning tasks / allocating resources for the tasks and monitoring that the mission is completed. The discussion was intended to challenge contemporary frameworks for military decision-making including a human-centric OODA-loop. This thesis has an AI-centric approach to the OODA loop and the joint targeting process. It could be criticized for adding incentives for a progression in decision support systems that takes the human out of the loop. This is alleviated by the scope of the research and the research gap within War Studies. Moreover, it is likely that Boyd would have approved or even encouraged deviations and adjustments from his own concept. After all, the nature of knowledge is a constant discourse or a battle for survival within an ever changing environment (Boyd 1976). The nature of war is also constantly changing. Neither force employment, nor processes or C2 will ever have a constant formation. As an example of this, the

predictive analysis technique employed in Model 1 contributes to a new way of understanding the "Observe" stage in Boyd's OODA loop and sets the stage for continued theoretical discussions. So does the tensions between speed and effectiveness in dynamic targeting or the 242 | EXPLORING DECISION ADVANT AGES theoretical trade-off between speed and depth as these relate to broader military theories on decision-making. The implications of the project towards ethics Although this applied research project was tailored by its research question, minor ethical considerations have been addressed within the thesis. With respect to ethical concerns of practicing research, the design, method and data collection has been planned, employed and conducted with little, if any ethical concerns while practicing the research (Guillemin and Gillam 2004; Ben-Ari 2014). Therefore, this section is dedicated to discussing ethical concerns to the employment of AI in the military domain

exposed through other researchers that are focused on ethics. Ethics are important aspects of decision-making and especially so if related to targeting. The implications of this project may contribute to the inclusion of AI for military organizations. It is therefore judicious to discuss how this study can have ethical implications. Some of the persistent ethical debates relate to the morality of drone warfare (Welsh 2015), the ethics of robotic warfare (Horowitz 2016) and turning warfare into an algorithmic warfare (Bode et al. 2024) As indicated throughout the literature review the current debated also revolve around the definition of meaningful human control (Steen et al. 2023; Verdiesen, Santoni De Sio, and Dignum 2021; J Johnson 2022b) This debate has been illustrated by researchers Ingvild Bode and Tom Watts in a recent article called Meaningless human control to highlight not only the importance of the term but also the blurriness within its concept formation (2021).

Nonetheless, this is not just a question of applying ethical concerns on the use of autonomous weapons or sensor systems at the tactical edge. Ethical considerations must be applied for the employment of AI as a means of intelligent automation and as intelligent agents within decision-making, be it in support of the ODDA-loop or in the joint targeting processes. Moreover, ethical concerns have to do with the whole spectrum, from research and development to producers and users. As this has become clearer and as numerous states, companies and universities are engaged in research, producing, and employing AI systems, the previous scarcity of governing frameworks is now slowly being addressed. The US Department of Defense adopted five ethical principles for AI development in 2020 (US DoD 2020) after which other states and organizations have followed up with their corresponding guidelines (Dunnmon et al. 2021) and a joint initiative called Responsible AI in the Military Domain(REAIM) (Delft

2023). Returning back to the US DoD original five ethical principles from 2020, they encompass five major areas: (1) “Responsible to the development, deployment, and use of AI capabilities”; (2) “Equitable to minimize unintended bias in AI capabilities”; EXPLORING DECISION ADVANT AGES | 243 (3) “Traceable to ensure appropriate understanding of the technology”; (4) “Reliable which include well‐defined uses, safety, testing and assurance”; and, (5) “Governable, to ensure AI’s intended functions while possessing the ability to detect and to disengage or deactivate deployed systems” (US DoD 2020, 1). Notwithstanding these ethical principles, it is the employment of AI, regardless of where, in the military domain, that is of importance from an ethical perspective. As Bode and Huelss argue in relation to the practices of autonomous weapon systems (AWS), that “AWS may shape norms in practice by privileging procedural norms that are detached from deliberative

processes.” (2021, 395) Bode and Huelss attempt to go beyond examining how AWS are governed by norms towards how the practices of using AWS create, shape, and define norms. This project has operationalized the basic idea behind ethical principles in research and development by employing transparency, accessibility and explainability in the experiments. These three aspects can be linked to the design and method employed Moreover, the two AI-models and the way in which they are theoretically applied and employed can support successive research on ethics to produce frameworks of policies and regulations for the use of AI in targeting. As evident from the extensive literature review, AI integrations in support of military decision-making is in its formative stage. However, tangible examples of what these AI-applications do and how they can be applied are uncommon and make this study a valuable illustration of ethical reflections. Accessibility and transparency were important aspects

during the design of this research. To enable ethical reflections from experts the project had to be conditioned in such a way that the problem could be researched in an open access environment. Arguably, research on warfare can be sensitive, especially if one imposes new ideas on warfare efficiencies or improved approaches explaining victory in battles as in Biddle’s study (2010). Applied research on AI in the military domain has been a reality for some time (Hoadley and Lucas 2018). It is this researcher’s true belief that the outcome will address benefits from integrating AI into joint targeting. Targeting practices would arguably become more efficient and effective if the results and conclusion of this research were to be incorporated within new targeting concepts. It could provide more precision, faster cycles of understanding, deciding, and acting, improved use of databases and more efficient sensor allocations. It could even be argued that Model 2 supports a more neutral

position to targets and targeting related thresholds and enables implementations of targeting guidance as decision policies in a multicriteria rule-based model. This argument also makes for a transition to the more practical contributions of the study. 244 | EXPLORING DECISION ADVANTAGES Practical contributions: Providing new means and approaches for practitioners. In the case of Model 1, predictive analyses and foreknowledge of high value targets (HVTs) are the core benefits accompanied by its utility for initiating target engagements. The findings show how deep learning can be applied for predictive analyses of adversaries’ valuable weapon systems, and through these estimates, a joint force can more efficiently allocate sensors to find and fix high-value targets. Pending a joint task force (JTF) targeting policies, the solutions from Model 1 could be utilized for tasking an air strike, given that the strike package can perform its own positive identification (PID) and field

collateral damage estimation (CDE). This approach would markedly shorten the time laps and could be applied to both existing methods (deliberate and dynamic) in joint targeting. Furthermore, Model 1 and its utilization of deep learning is a precursor of how neural networks can contribute to what is referred to as intelligence support to targeting. Performing estimates and producing foreknowledge with deep learning techniques is likely to expand into other areas such as wargaming during operations and building situational awareness on data retrieved from missions within an operation as well as post-strike activities like combat assessments (CA). CA and managing the battlespace also represent the foundation for associated combat engagements and deliberate targeting activities. In the case of Model 2, speed is the core benefit along with the ability to incorporate targeting policies into this multi-criteria rule-based application. The findings show how machine reasoning using optimization

algorithms can produce target engagement options (attack options), augmenting decision-making in performing swift target engagements on prioritized targets of opportunity. Emergent target can be immediately acted against using an application of machine reasoning to optimize parallel target engagements. In this experiment 16 parallel dynamic target engagements were conducted using 13 different joint weapons, which enabled over 330 different options pending factors that included distance to targets, munition feasibility, target vulnerability, and probability of kill. Notwithstanding that optimization is a rule-based AI approach, mathematical optimization is a powerful AI technique that seeks optimum decisions and outcomes of a given problem. An essential characteristic of an optimization problem is that it involves a set of variables that interact in complex ways. It is beyond human limits to keep track of all the interactions and potential outcomes in even relatively simple optimization

problems as Model 2 represents. Certainly, maximizing targeting outcomes or in this case configured through a multicriteria function and given eight constraints can be used for other problems within a targeting enterprise. Moreover, equations are generic representations and can be optimized repeatedly using diverse and countless data sets as well as alternative objectives. Each result would then represent a best-case scenario that can be EXPLORING DECISION ADVANT AGES | 245 evaluated and applied. In the case of recommending attack options, Model 2 could be duplicated and primed with different data sets or objectives to generate more options for the decision-maker in which Option A is primed to represent one targeting approach and where Option B, C and n all have other configurations. Optimization models allow for tailored decision policies being executed exactly and consistently every single time, given that it relies on an exact method as opposed to a heuristic method. If the

model receives preprocessed information about a target (positive ID and target coordinates) by connected sensors and similarly be updated by connected weapon systems this would create a web of targeting services for decision-makers. It could become a vital planning tool to reduce the total time of target engagement. Given the level of confidence and trust that a human decisionmaker would have in this model, the output could initiate a strike assignment Collectively, the two AI-applications provide new means and approaches to dynamic targeting for practitioners. They expand a commander’s decision-making opportunities. The simulations showcase how a commander's targeting guidance can be utilized as multi-criteria rules that can be implemented into the targeting recommendations ensuring that each recommendation has considered all given aspects including relevant constraints mathematically. The key findings and the practical contribution from the two experiments have immediate

relevance to military organizations. The study shows that dynamic targeting can be orchestrated in time, space and with forces. This opens for a transition from hierarchical targeting structures to more dynamic and non-linear concepts where synchronization issues could be mitigated by implementing targeting web services. This is further elaborated within the next section Practical implications: Expanding the decision space This section focuses on the operational and organizational opportunities and challenges of integrating AI into dynamic targeting. It ties these themes to real-world challenges and solutions, focusing on how AI can be used to adapt workflows and C2 structures for decision-making in a joint targeting context. The results indicate that both AI-models improve the ability to make informed decisions. The use of data as the source and AI as the interpreter of it creates outcomes that expand the decision space. The term is used here to capture the expansion of options

provided by the models in relation to if they had not been employed. The point is that if the output of these AI systems matter and add value to a decision-maker and that this otherwise would have been difficult or even impossible to achieve, then the decision space has expanded. It could be time-related as in Model 2 where this would be possible by humans, though with a massive delay (one second compared to >10 minutes). It could also be related to predictions from 246 | EXPLORING DECISION ADVANT AGES immense computations that are seemingly unobtainable as in Model 1, where AI learns the signal attenuation effects of topography at a pixel level from satellite images. The integration of AI for decision support is already under way and makes it a critical military research area. New AI platforms (AIPs) are likely to soon provide automation and autonomy in battlespace management systems (BMS). These offer an understanding of a given situation that is even deeper and broader at

all levels of warfare. Although it would prerequisite the sharing of data in-between more entities within a system, it could theoretically expand the decision space at more than one command simultaneously. Just as the two models are foreseen to be integrated, other AIPs could also be constantly updated and display the latest and most relevant information or knowledge for anyone requiring it (given their access). Such AIPs would perform constant iterations of collecting-filtering-presenting, conditioned by their settings, and subject to subscription of it. However, all accessible data and information is not necessarily relevant. Filtering will be vital in the context of expanded decision spaces. Furthermore, the ability for AI to manage uncertainties is limited as of now. Consequently, the risks associated with machines taking over the processes that lead to decision-making by collecting-filtering-presenting what it assumes is valuable information must be governed by control mechanisms

that have as a minimum a probability of certainty of the information presented by the system. This could be perceived as a synthetic peer-review of the content. Importantly, the uncertainties inherent in warfare may become less apparent to humans as more AIPs enter the decision calculus. At least until more credible context-aware AIPs become operational. Context-aware AI will have the ability to tweak the current Observe-Orient towards sense-compare-explore-update-forecastpredict (or prescribe). Advanced versions will arguably soon be able to collect data as it sees fitting the overarching objective. Any associated uncertainties will be detected and either managed by additional collection taskings or accepted with a computed risk attached. Moreover, such a system will likely explore, learn, and constantly update its internal belief system with external data, whilst managing an oversight of updates to avoid deception and misinformation. The key takeaway from this implication is that

there are opportunities to expand the decision space and create more options that allows adaptation to different scenarios or contingencies. This includes options to avoid becoming predictable Overall, the practical utility of expanding the decision space is that it can enhance operational efficiency and decision-making. EXPLORING DECISION ADVANT AGES | 247 Towards novel targeting practices The thesis suggests that current targeting practices will transition towards new targeting concepts. Modelling and simulation environments where AI-applications can be explored are becoming a breeding place for novel abilities, capacities, and employments. The project’s two experiments were data-driven and exploited AI’s abilities in two different problem settings. It enabled a pathway to novel insights on AI augmentation to targeting decision-making. Arguably, the result from the empirical experimentation suggests that Model 1 and Model 2 are representative of AI-based targeting ‘web

services’ that can offer decision advantages. Similar web services could become the centerpiece in future targeting concepts. It could be a top node or hub that benefits DARPA’s Mosaic Warfare ideas. Their Strategic Technology Office (STO) describes these as distributed “sense-decide-and-act systems across a wide number of platforms [that] mass your firepower without having to mass your forces” (2018b, 1). Bryan Clark has referred to it as exploiting AI to create decision-centric operations (2019). Nevertheless, as evident in the second experiment, the model creates over 300 feasible attack options given its input of data. This is powerful, fast and precise. If the model had been connected to a network of sensors and effectors and accessible to more than one decision-maker, then there be more options allowing more asynchronous actions. It would improve the “capacity for independent action” as Boyd suggests (Osinga 2007, 132). Furthermore, AI applications can co-create

robust systems if attainable by every actor with decision authority in a network. Boyd discusses these matter in a number of ways to include in relation to self-organization (organic, adaptive and cellular), characterizing these as “chains of interdependency branch in complicated patterns across nearly every actor in a broad network of interaction” (Ibid. 2007, 114) and in relation to improving the capacity for decision and action by enabling subsystems to operate “with a sufficient level of autonomy, yet in an interdependent way, with other systems” (Ibid. 2007, 124). Finally with regard to the key theme of adaptation, implicit control, shared awareness and the idea that “initiative will not be fostered by top-down, directive command and control systems.” (Ibid 2007, 199) These examples also circle back to Boyd’s early conception of operating at a faster tempo or rhythm than the adversary. The ‘rapid OODA loop’ correlate more to decision-making within a dynamic

targeting context than his later conception of it as a strategy. Furthermore, other ideas within his strategy do include the notion that novelty matters: “by reorienting ourselves to match a changing world affected by science, engineering and technology [] creates opportunities for novel concepts to arise from new insights, imagination and initiative” (Osinga 2007, 228). The study and its empirical investigation are arguably artefacts from an approach similar to Boyd’s conceptual spiral where the two AI-applications can be perceived as forerunners of AI-based targeting web services. 248 | EXPLORING DECISION ADVANT AGES Strategic contributions: The importance of speed and the necessity of leveraging it One of the most significant findings discussed throughout the thesis is speed. The findings on the matter offer policy-level insights suggesting shifts in organizational practices and defense modernization strategies and affect command structures. Targeting relates to time in

several ways and takes its form in different targeting contexts. The thesis argues that given the amount of data required to inform humans prior to targeting decisions (or actions), AI is needed to augment this to various degrees (Table 6). The more AI that is integrated into a targeting enterprise (including AI-enabled weapon and sensor systems at the tactical edge) the more relevant time (speed) becomes for the processes within targeting cycles. With the assumption that AI can produce credible predictions on high value target locations, it will make the sensor allocation and the intelligence collection more efficient creating more opportunities and target engagements. Furthermore, accepting that more dynamic opportunities arise from detecting and managing emergent targets through more efficient battlespace management systems, the attempt to engage those will impose accelerated targeting cycles and decision-makings. As established in the second experiment, rule-based models can

expedite immediate and feasible recommendations for target engagement. These recommendations can be turned into tasking orders to effectors within seconds after decisions are made if integrated into a larger system where these effectors are interconnected. At the most abstract level speed relates to being faster than the adversary. Given a more tangible form, it relates to the window of vulnerability that can be attributed to all targets, as well as the window of opportunity if referring to the availability of effectors to engage during each target’s window of vulnerability. This will require rapidity and agility from AI-enabled targeting web services within a military organization and its targeting concepts. A Targeting Nexus (TN) could offer a solution to this. EXPLORING DECISION ADVANT AGES | 249 Strategic Implications Future direction 1: The broader consequences for military policy and strategy by moving towards a Targeting Nexus The broader strategic implications pertain to

the potential to build upon this study and its findings. It begins by establishing the Targeting Nexus (TN) as a concept to describe an AI-enabled system designed to integrate and enhance the military targeting process. The Targeting Nexus serves as the core function for synchronizing data-driven insights, human decision-making, and operational execution across the sense-decide-act continuum. By combining advanced analytics, AI-enabled web services and human oversight, the (TN) provides a cohesive framework for managing the complexities of modern targeting operations. It emphasizes adaptability, precision, and the integration of human and machine intelligence capabilities to optimize targeting outcomes while maintaining accountability and operational control. Consider in this light a Joint Task Force (JTF) with multiple targeting directorates operating within a unified Targeting Nexus (TN), where all resources are interconnected and accessible to every directorate. Figure 38

illustrates this baseline configuration of the TN, showing the integration of targeting directorates and resources. Targeting directorates On-call resource layer Pre-tasked resource layer Active resource layer Inaccessible resource layer Figure 38 Illustration of the baseline configuration of multiple targeting directorates and all interconnected targeting resources within a random JTF organization. The design is rooted in its inherently digitalized components, which are viewed as the most logical structure given the opportunities provided by digitalization (Derian and Rollo 2023). Within this framework, the targeting directorates operate at the same level of command and authority. The Targeting Nexus facilitates critical processes such as target vetting, validation, and other functions essential for maintaining fluidity and adaptability to rapid sense-decide-act cycles. As illustrated in Figure 38, the On-call resource layer encompasses all currently available targeting resources

(TRs) assigned to specific targeting directorates (TDs). However, not all 250 | EXPLORING DECISION ADVANT AGES resources are accessible to all directorates due to factors such as geographical constraints or operational limitations. The Pre-tasked resource layer includes TRs already assigned to a TD but not yet engaged in their mission. In contrast, the Active resource layer comprises TRs currently executing their assigned tasks. Finally, the Inaccessible resource layer represents resources unavailable to all TDs due to maintenance, munition shortages, or other restrictions. A rule-based application would oversee the operation of a flat C2 structure with multiple TDs, ensuring adherence to established rules and policies across all targeting web services. Supervision of the Targeting Nexus would be activated when two or more directorates request the same resource. This supervisory layer is depicted alongside examples of web services in Figure 39. AI-based Targeting Web Services

Supervisor for Assignment Problems Dynamic Targeting Targeting Policies Collection Operations Management Collection Policies Figure 39 Visualization of the AI-based targeting services that could be provided within a targeting enterprise. One method for supervising the allocation of targeting resources (TRs) is the auction algorithm (Bertsekas 2008),which optimizes the assignment problem to maximize the total benefit for all bidders (targeting directorates). Alternatively, a global optimum could be achieved by adapting Model 2, modifying the decision variables, objective functions, or constraints to suit the specific optimization problem. This approach could incorporate existing resource prioritization procedures used in contemporary joint targeting (see Fig. 8 in Chapter 3) These procedures consider factors such as (1) the threat severity posed by the target to the JTF, (2) the target's EXPLORING DECISION ADVANT AGES | 251 strategic significance based on its value to the

adversary or the JTF, and (3) time sensitivity, including the target's threat window and the JTF’s opportunity to act 103. Incorporating a cost-benefit analysis into the supervisory framework could enhance its operational value. For instance, matching the cost of munitions to the estimated value of the target would enable a return on investment (ROI) calculation for each Targeting Directorate’s request for targeting resources. This criterion would operationalize the principle of economy of force, similar to the approach employed in Model 2. Broader performance metrics could further refine the system, including timeliness in response to requests, accuracy of information, accounting for system latency and strategies to address ‘age-of-information’ (AoI) challenges (Yang et al. 2023, 2), computational reliability, and service consistency. Yang et al defines Age-ofInformation in their work as “the time elapsed from the moment when the information was prepared by the end

devices to the moment when the information arrived at the destination” (2023, 2). In most cases, one would strive to minimize the AoI, as the lower AoI the higher the efficiency of an information scheduling process. Establishing a Targeting Nexus to oversee multiple targeting directorates would facilitate simultaneous decision-making, allowing decision-makers to capitalize on emerging opportunities within their respective battlespaces. Tactical commanders across the targeting enterprise could act independently while adhering to a unified and dynamically updated targeting decision policy. Individual sensors or effectors could actively contribute by monitoring their surroundings and communicating their capabilities and readiness (e.g, ‘I am here and capable of performing this action now’) to the Targeting Nexus. If the JTF’s intelligence management processes were to be integrated an AIP could autonomously task or retask ISR platforms to designated target areas of interest,

supporting advanced target development. Additionally, the system could manage intelligence collection efforts more effectively, coordinating and synchronizing assets while addressing deconfliction requirements. This integration would enable a seamless flow of actionable intelligence and optimize resource allocation within the targeting enterprise. 103 This can also be referred to as window of vulnerability. 252 | EXPLORING DECISION ADVANT AGES Future direction 2: The strategic implications of context-aware AI The next stage in this research vector is context-aware AI that is likely to transform executive roles within military operations. While not part of the two experiments conducted in this thesis, a prototype was developed as a spin-off within the doctoral studies based on Model 2. This prototype demonstrated how context-aware AI could be integrated into targeting-as-a-service, providing a proof-of-concept for future developments. Biddle emphasized in 2010 that “effective

force employment” (defined as “the doctrine and tactics for using material resources”) was central to warfighting yet often overshadowed by “technological modernization” (2010, 2–4). While Biddle’s attention to readiness and training remains valid, the integration of AI shifts the conversation to how both material and non-material (AI) are employed. Findings from this thesis demonstrate that AI can augment targeting decisions in ways beyond human limits, enhancing the efficiency and speed of dynamic targeting while laying the groundwork for AI-enabled targeting web services and advanced Targeting Nexus constructs. Moreover, training and readiness are likely to be facilitated through modern simulation environments, such as defense metaverses, where context-aware AI systems can be effectively developed and tested (Borchert et al. 2022) Warfare evolves through technologies as shown by current conflicts in Ukraine and Gaza. The integration of AI demands innovative solutions

and moves beyond traditional paradigms to adapt concepts to new realities. Context-aware AI systems exemplify this shift and indicate an increasing role for intelligent systems in military decision-making and potentially autonomous execution. The strategic implications for command and control (C2) structures are profound. Advances in generative AI (GenAI) and multi-agent systems (MAS) enable autonomous agents to perceive, decide and act collaboratively in shared environments (Nguyen, Nguyen, and Nahavandi 2020; W. Wang et al 2024) These context-aware multi-agent systems (CA-MAS) adapt dynamically to changes in the environment and other agents' behaviors offering real-time coordination capabilities as demonstrated in swarms of autonomous drones allowing them to self-organize and execute the given tasks (Borchert et al. 2022) In this thesis, Model 1 and 2 demonstrated similar but separate capabilities with Model 1 employing deep neural networks for prediction and sensor allocation

and Model 2 using optimization algorithms to achieve the best solution possible. If combined, they would form a basic MAS. Furthermore, if given the authority, the output of Model 1 could be inputs for the collection operations management within the intelligence process termed TCPED 104 (UK Ministry of Defense 2023, 4) as well as inputs to Model 2 that could compute recommendation and assign the best 104 Task-Collect-Process-Exploit and Disseminate. EXPLORING DECISION ADVANT AGES | 253 solution for target engagement. However, it would still require a context-aware central planning resource. A balanced approach could involve maintaining human decision-makers at the core while integrating context-aware AI as an additional feature in the Targeting Nexus. By incorporating an intermediary layer of generative AI, this setup would enable interactions with context-aware agents, ensuring explainability of datasets and solutions. This concept was prototyped for the Swedish Joint

Headquarters in October 2024, using Model 2 as the foundation. Key features of this research prototype are illustrated in Figure 41. GenAI process: PERCEIVE-REASON-ACT-LEARN AI-based Targeting Web Services Dynamic Targeting Dataset Rule-based Optimizer Targeting database Targeting Policies Targeting policies GenAI Context awareness Context-aware reasoning Memory Targeting Directorate Context awareness LLM ‘World Model’ Situation or scenario Environment ‘Real world’ GenAI agent Optimization Model Figure 39 A simplified model of the MVP that was demonstrated at the Swedish Joint Headquarters in October 2024. The figure above illustrates the role of the generative AI (GenAI) agent as an intermediary in the interaction between human operators within a Targeting Directorate (TD) and the rule-based Optimizer. Acting as both a chatbot and a manager, the GenAI agent facilitates communication and decision-making by leveraging a large language model (LLM) as its

foundational framework, incorporating a targeting ontology and relevant expressions. The process begins with the TD representatives identifying themselves, allowing the GenAI agent to recognize participants and accurately record discussions and decisions. Through speech-to-text functionality, TD members can communicate directly with the GenAI agent, which interprets the input and interfaces with the Optimizer to provide actionable options. This ensures that both the TD and the GenAI agent maintain situational awareness (or scenario understanding in training contexts). The GenAI agent records all interactions, translating speech into text and storing each targeting engagement in a memory database. This memory can be referenced by the TD for future decision-making and by the GenAI agent for learning purposes. Importantly, it would enable transparency and explainability as well as clarifications and justifications to decisions made in the past. 254 | EXPLORING DECISION ADVANT AGES

The continued evolution of context-aware AI systems will redefine military C2 structures, enabling more efficient, adaptive, and explainable decision-making processes while maintaining human oversight where needed. Limitations and reflections Technological and methodological limitations The two models developed and tested in this study are not without limitations both technologically and methodologically. Technologically, the models represent specialized AI that is designed to solve specific problem sets within their respective domains. Methodologically, the research relied on artificial constructs and controlled experimental conditions that limit the generalizability of the findings beyond these settings. The results are inherently tied to the specific contexts and assumptions under which the interventions were conducted. Consequently, their applicability is restricted to these predefined scenarios and cannot be readily generalized to broader contexts without adjustments. Despite

these limitations, the models offer flexibility for adaptation to other situations or tasks. For instance, Model 1, which predicts optimal radar locations based on topographic and RF signal data, could be extended to locate any ground-based surface-to-air missile system, provided the input parameters align with the terrain and operational environment in question. Similarly, Model 2, focused on optimization algorithms for targeting decisions, could be adjusted to incorporate different targeting directions, rules of engagement or policy constraints, transforming the objective functions to fit diverse operational needs. Simulated environment and human interaction Both models were tested in simulated environments without direct interaction with human operators. Inputs to the AI were structured as predefined problem statements, effectively acting as a proxy for human requests. While the experiments demonstrate the potential of the models to function as part of a decision-making system, the

absence of real-time human-AI interaction introduces a gap in understanding how these systems would perform in dynamic, real-world settings. Future research should focus on integrating these models into live exercises or operational environments to evaluate their efficacy and usability when directly interfacing with human decision-makers. EXPLORING DECISION ADVANT AGES | 255 Scope and experimental design The scope of this study focused on two critical components of the targeting processintelligence and operations represented by Models 1 and 2 respectively. These models were designed as applied AI platforms (AIPs) that could potentially be integrated into larger systems or networks. While the study demonstrated their potential to augment specific targeting tasks, it did not encompass the broader decision-making continuum, such as deliberate joint targeting or tactical-level combat engagement. The limited scope leave open questions about how these models would perform if scaled to

encompass the full range of targeting activities. The experimental design also faced constraints related to data and scenario complexity. Simulated environments allow for controlled experimentation but do not fully replicate the uncertainties or operational constraints present in real-world military contexts. Moreover, the models’ reliance on structured data inputs underscores the need for further research into how AI systems can process and adapt to incomplete or even ambiguous data. Applicability to real-world contexts The findings demonstrate the potential for AI but their applicability to real-world contexts remain constrained by the artificial nature of the experimental environment. Real-world military operations involve complex challenges that include coordination across different levels of command, competing priorities and counteractive measures from an adversary. Integrating the proposed models into real-world scenarios would require addressing these complexities, as well

as considerations of trust, transparency, and accountability in human-AI collaboration. Additionally, implementing AI systems at scale would necessitate overcoming barriers within military organizations, as well as ensuring compliance with ethical and legal frameworks. This study faced several challenges that included the complexity of modelling and ensuring the accuracy and relevance of data and verifying / validating the results of the simulations. Although the aim and objectives of this project have been fulfilled, it has highlighted some areas for future investigation within the next section. 256 | EXPLORING DECISION ADVANTAGES Directions for future research. This section briefly explores some of the unresolved issues and areas for further exploration. The empirics have shown two roles where AI can augment human decision-making within dynamic targeting. It seems likely that military organizations will have an increased acceptance towards more automation and autonomy. They will

incrementally acknowledge conditions and circumstances where they cannot afford to have a human in the loop or even on the loop (Scharre 2018a). How should new targeting concepts be configured to meet requirements, for instance, meaningful human control (Bode and Watts 2021), whilst introducing new capabilities and responsibilities enabled through integrations of progressively more intelligent agents? Transitions of roles and responsibilities will affect the structure, the processes and methods within military organizations. It will affect the strategic culture within and between organizations. If a military organization strives to be active rather than reactive against a peer adversary and adaptive to the changing operational environment where it is to “compress own time and stretch-out adversary time” (Osinga 2007, 141–42), then it needs to reconfigure its joint targeting concepts. Decades ago, Van Creveld suggested five interacting implications for the organization of command

systems and how they operate, including, the ‘need for decision threshold’ to be fixed as far down as possible, provision for ‘self-contained units’, and ‘information-transmission throughput’ in all directions(1985, 269–70). Arguably, this was a postulation of networked, adaptive organizations based on historical reflections of command. It resembles Boyd’s idea that ensuring the proficiency of compressed OODA loops at the higher levels of command produces exponential gains to the OODA loops at lower levels. The thesis has suggested implications that support these notions on C2 However, they need to be verified and validated in more realistic settings. Human-AI-integrations including the anticipated command-and-control reconfigurations are yet to be resolved. They fuse two key aspects together in the form of ethics and trust. Unlike the former, which is one-directional and in this context refers to the human employment of AI, the latter is bi-directional in the sense

that humans must place a level of trust in the AI systems they integrate into its processes, staff work, and decision-making. The thesis offers some direction of applied philosophical research on ethics and trust. In the book Call Sign Chaos, former US Secretary of Defense General (ret.) James Mattis states that “operations occur at the speed of trust” (2019, 156) referencing the relationship between human commanders and their subordinates. Arguably, this is true also in the case of commanders / staff and their forthcoming relations with AI systems. Put simply, without trust, these AI systems will not be employed. Additionally, redistributions of roles and responsibilities within a targeting organization require ethical considerations and transparent states of accountability. EXPLORING DECISION ADVANT AGES | 257 The current approach to how humans can measure their trust in AI is in AI’s performance which relates to experiences. The more experience that humans have with AI

performing satisfactorily, the higher the trust. This will in turn lead to an expectation and a sense of reliability, but not necessarily transferred to humans’ trust in other AI systems. As the understanding of AI increases, the likelihood of adapting ethically accepted and trusted AI platforms in different decision-support systems rises. Boyd saw adaptation as a matter of life and death because to adapt was to survive. Properly adapting to a growing number of smarter sensors, weapons and AI applications that can optimize the use of resources in dynamic scenarios is likely to become a core capability of military organizations. Human-machine integration is also a question of calibrating human control. David Mindell have suggested that it “takes more sophisticated technology to keep the humans in the loop than it does to automate them out” (2017, 56). Any calibration will therefore depend on numerous factors including the complexity of the task, risks and consequences involved.

Experiments with staff from the Marine Corps Tactics and Operations Group have shown that commercial of-the-shelf war gaming engines can be used to for experiments for instance observation of human-machine collaboration, as a forum to understand how humans balance data and intuition under stress and uncertainty and to find the right balance of human-machine in future war (Jensen, Cuomo and Whyte, 2018). There may also be situations where not having a human in the loop is unacceptable no matter what. Moreover, future research should explore how human operators collaborate with AI systems in real-time, particularly in high-pressure, dynamic operational contexts. This includes understanding how trust, transparency, and explainability can be enhanced to foster effective collaboration. Expanding this project to cover the full targeting continuum from pre-planned joint targeting at the strategic level to real-time combat engagements at the tactical level would provide a more comprehensive

understanding of AI’s potential impact on military organizations. Other areas that are foreseen to be of importance are investigating the implications of AI-driven decision-making on accountability, responsibility, and compliance with international laws and norms is critical as these systems move closer to operational deployment. Moreover, addressing issues related to data availability, quality and integrity in operational environments remains a key area for future work. Research should focus on developing AI systems that can function effectively under conditions of uncertainty and incomplete information. 258 | EXPLORING DECISION ADVANT AGES Concluding thoughts The research conducted in this study demonstrates the potential for AI to augment critical components of the military targeting process, but it also underscores significant constraints and unresolved questions. The central research question to this project was: how can AI augment human decision-making to create decision

advantages within dynamic targeting settings? The key findings suggest that AI can augment human decision-making using neural networks and multi-criteria rulebased algorithms to improve precision and efficiency in sensor allocations and to recommend optimal attack options. EXPLORING DECISION ADVANT AGES | 259 Key terms and definitions Actor – human or non-human entities or agents with abilities that can perform actions and participate within a system or network for specific purposes. Artificial Intelligence (AI) – “human-like intelligence displayed by software and/or machines” (Hillier and Lieberman 2021, 21) Command and Control (C2) - US DOD dictionary defined C2 as “The exercise of authority and direction by a properly designated commander over assigned and attached forces in the accomplishment of the mission.”(Office of the Chairman of the Joint Chiefs of Staff 2020, 40) Comprehension – a factor used in the project defined as an ability to absorb and make sense

of data and/or information. It also refers to capability limits of human cognition Intelligent agents – “something that perceives and acts" (Russell and Norvig 2014, 34). The intelligent agent is defined by three factors: “the nature of the environment in which the agent operates; the agent's connections to the environment; and the agent's objectives” Russell (2019, 43). Joint targeting process – See military targeting. The joint targeting process resides at the operational level of command. Knowledge – a socially generated product based on an interplay between actors to attain certain goals or desirable situations. Machine Learning (ML) – A subset of artificial intelligence, where the machine learns from being exposed to data in a training environment for specific objectives. Deep learning is a subset of machine learning, and a more sophisticated method of learning, where the artificial system is opting itself. (Lecun, Bengio, and Hinton 2015). Put

more simply, algorithms that “can learn from data to make predictions,” and by doing so progressively improve its performance on a specific task (Hillier and Lieberman 2021, 21) Military targeting – the process, or cycle, of selecting and prioritizing targets and matching the appropriate response to them, considering operational requirements and capabilities (US DoD Joint Publication 2018b). Furthermore, NATO define the this targeting cycle as linking “translates strategic-level direction and guidance and the Commander JTF’s direction and guidance at the operational level into tactical level activities that service targeting priorities.” (NSO 2021, Edition B:1–13) Modelling and Simulation (M&S) – A scientific method characterized by first understanding the problem by gathering all relevant data, then constructing a model in an attempt to abstract the essence to represent the essential features of the 260 | EXPLORING DECISION ADVANT AGES situation or conditions,

so that the conclusions (or results) obtained from the model are valid also for the real problem. Speed - a factor (of time) that, in this thesis, sets the limits of the time available to perform decision-making Targeting enterprise – Includes the forces, assets, processes, practices and C2organization used for targeting within a Joint Force structure. Targeting Nexus (TN) – A new concept suggested to denote and describe an AIenabled system designed to integrate and enhance the military targeting process. A core function for synchronizing data-driven insights, human decision-making, and operational execution across the sense-decide-act continuum. By combining advanced analytics, AI-enabled web services and human oversight, the (TN) provides a cohesive framework for managing the complexities of modern targeting operations. It emphasizes adaptability, precision, and the integration of human and machine intelligence capabilities to optimize targeting outcomes while maintaining

accountability and operational control. EXPLORING DECISION ADVANT AGES | 261 Appendix A - Survey 2023 as part of the evaluation of Model 1 262 | EXPLORING DECISION ADVANT AGES EXPLORING DECISION ADVANT AGES | 263 264 | EXPLORING DECISION ADVANT AGES Appendix B - Example of Model 2 scripting log Version identifier: 22.110 | 2022-11-28 | 9160aff4d Legacy callback pi Multi-objective solve log . Index Priority Blend Objective 1 4 1 5.0000000000e+00 2 3 1 4.0000000000e+00 3 2 1 1.6000000000e+01 4 1 1 -7.6550000000e+06 5 0 1 1.3000000000e+01 Multi-objective scripting log Nodes Time (sec.) DetTime (ticks) 0 0.01 6.04 0 0.01 7.34 0 0.00 6.40 0 0.00 3.57 0 0.00 2.83 T1HVT1C4IHQ Large (BrigCP)742 T2CCT3C4IHQ Medium (BatCP)422 T3HVT1ADFSAM Medium Range1050 T4CCT3ADFSAM Short Range1050 T5TST1ADFSAM Long Range1052 T6CCT3AFAAirstrip211 T7HVT2GFFArtillery Unit Medium Range330 T8CCT3GFFArtillery Unit Short Range330 T9HVT2AFFAttack Helo Site662 T10CCT3GFFTank Unit330

T11TST1MSLSS Missile Long Range1050 T12HVT2MSSLogistics depot Large442 T13CCT3MSSLogistics depot Medium222 T14HVT1LOCBridge Large332 T15HVT2LOCBridge Medium222 T16CCT3LOCBridge Small111 E1HowitzerArcher HE 77 BB2500001033NoStationary1000 E2HowitzerArcher Bonus300001033NoStationary25000 E3HowitzerArcher Excalibur30001045NoStationary110000 E4HIMARSATACMS MGM 140B5050300NoStationary850000 E5GMLRSGLSDB (GBU-39)14050150NoStationary40000 E6AircraftGBU-12 300115YesStationary22000 E7AircraftGBU-49150115YesMoving43000 E8AircraftKEPD-350301400YesStationary1090000 E9AircraftAGM-88E301150YesStationary870000 EXPLORING DECISION ADVANT AGES | 265 E10DroneTB2 MAM-L180115YesStationary25000 E11DroneMQ-9 w GBU-12300115YesStationary22000 E12DroneMQ-9 w GBU-49200115YesMoving43000 E13DroneMQ-9 w AGM-114 Hellfire II15018YesMoving150000 A001T1HQ Large (BrigCP)E8KEPD-35010142750042750NoYesYes A002T1HQ Large (BrigCP)E8KEPD-350112100500070500YesYesYes A003T1HQ Large

(BrigCP)E8KEPD-3501231432500113250YesYesYes A004T1HQ Large (BrigCP)E8KEPD-35020290000090000NoYesYes A005T1HQ Large (BrigCP)E4ATACMS MGM 140B10142750042750NoYesYes A006T1HQ Large (BrigCP)E4ATACMS MGM 140B112100500070500YesYesYes A007T1HQ Large (BrigCP)E4ATACMS MGM 140B1231432500113250YesYesYes A008T1HQ Large (BrigCP)E4ATACMS MGM 140B20290000090000NoYesYes A009T1HQ Large (BrigCP)E7GBU-4920240500040500NoYesYes A010T1HQ Large (BrigCP)E7GBU-4921375750045750YesYesYes A011T1HQ Large (BrigCP)E7GBU-4930364125064125NoYesYes A012T1HQ Large (BrigCP)E7GBU-49325121875091875YesYesYes A013T1HQ Large (BrigCP)E7GBU-4940490000090000NoYesYes A014T1HQ Large (BrigCP)E6GBU-1260643200043200NoYesYes A015T1HQ Large (BrigCP)E6GBU-1263985200049800YesYesYes A016T1HQ Large (BrigCP)E6GBU-1280864800064800NoYesYes A017T1HQ Large (BrigCP)E6GBU-128412112200082200YesYesYes A018T1HQ Large (BrigCP)E6GBU-121001090000090000NoYesYes A019T1HQ Large (BrigCP)E11MQ-9 w GBU-1260643200043200NoYesYes A020T1HQ Large (BrigCP)E11MQ-9 w

GBU-1263985200049800YesYesYes A021T1HQ Large (BrigCP)E11MQ-9 w GBU-1280864800064800NoYesYes A022T1HQ Large (BrigCP)E11MQ-9 w GBU-128412112200082200YesYesYes A023T1HQ Large (BrigCP)E11MQ-9 w GBU-121001090000090000NoYesYes A024T1HQ Large (BrigCP)E12MQ-9 w GBU-4920240500040500NoYesYes A025T1HQ Large (BrigCP)E12MQ-9 w GBU-4921375750045750YesYesYes A026T1HQ Large (BrigCP)E12MQ-9 w GBU-4930364125064125NoYesYes A027T1HQ Large (BrigCP)E12MQ-9 w GBU-49325121875091875YesYesYes A028T1HQ Large (BrigCP)E12MQ-9 w GBU-4940490000090000NoYesYes A029T1HQ Large (BrigCP)E5GLSDB (GBU-39)20240500040500NoYesYes A030T1HQ Large (BrigCP)E5GLSDB (GBU-39)21375750045750YesYesYes A031T1HQ Large (BrigCP)E5GLSDB (GBU-39)30364125064125NoYesYes A032T1HQ Large (BrigCP)E5GLSDB (GBU-39)325121875091875YesYesYes A033T1HQ Large (BrigCP)E5GLSDB (GBU-39)40490000090000NoYesYes A034T2HQ Medium (BatCP)E7GBU-4910142750042750NoYesYes A035T2HQ Medium (BatCP)E7GBU-49112100500070500YesYesYes A036T2HQ Medium

(BatCP)E7GBU-491231432500113250YesYesYes A037T2HQ Medium (BatCP)E7GBU-4920290000090000NoYesYes A038T2HQ Medium (BatCP)E6GBU-1230343740043740NoYesYes 266 | EXPLORING DECISION ADVANTAGES A039T2HQ Medium (BatCP)E6GBU-1231474940044940YesYesYes A040T2HQ Medium (BatCP)E6GBU-1240461560061560NoYesYes A041T2HQ Medium (BatCP)E6GBU-12437127860097860YesYesYes A042T2HQ Medium (BatCP)E6GBU-1250590000090000NoYesYes A043T2HQ Medium (BatCP)E11MQ-9 w GBU-1230343740043740NoYesYes A044T2HQ Medium (BatCP)E11MQ-9 w GBU-1231474940044940YesYesYes A045T2HQ Medium (BatCP)E11MQ-9 w GBU-1240461560061560NoYesYes A046T2HQ Medium (BatCP)E11MQ-9 w GBU-12437127860097860YesYesYes A047T2HQ Medium (BatCP)E11MQ-9 w GBU-1250590000090000NoYesYes A048T2HQ Medium (BatCP)E12MQ-9 w GBU-4910142750042750NoYesYes A049T2HQ Medium (BatCP)E12MQ-9 w GBU-49112100500070500YesYesYes A050T2HQ Medium (BatCP)E12MQ-9 w GBU-491231432500113250YesYesYes A051T2HQ Medium (BatCP)E12MQ-9 w GBU-4920290000090000NoYesYes A052T2HQ Medium

(BatCP)E5GLSDB (GBU-39)10142750042750NoYesYes A053T2HQ Medium (BatCP)E5GLSDB (GBU-39)112100500070500YesYesYes A054T2HQ Medium (BatCP)E5GLSDB (GBU-39)1231432500113250YesYesYes A055T2HQ Medium (BatCP)E5GLSDB (GBU-39)20290000090000NoYesYes A056T2HQ Medium (BatCP)E10TB2 MAM-L40436000036000NoYesYes A057T2HQ Medium (BatCP)E10TB2 MAM-L43778000048000YesYesYes A058T2HQ Medium (BatCP)E10TB2 MAM-L60660750060750NoYesYes A059T2HQ Medium (BatCP)E10TB2 MAM-L6410116250086250YesYesYes A060T2HQ Medium (BatCP)E10TB2 MAM-L80890000090000NoYesYes A061T2HQ Medium (BatCP)E3Archer Excalibur40454000054000NoYesYes A062T2HQ Medium (BatCP)E3Archer Excalibur41582500052500YesYesYes A063T2HQ Medium (BatCP)E3Archer Excalibur50571250071250NoYesYes A064T2HQ Medium (BatCP)E3Archer Excalibur527114750084750YesYesYes A065T2HQ Medium (BatCP)E3Archer Excalibur60690000090000NoYesYes A066T3SAM Medium RangeE9AGM-88E40454000054000NoYesYes A067T3SAM Medium RangeE9AGM-88E41582500052500YesYesYes A068T3SAM Medium

RangeE9AGM-88E50571250071250NoYesYes A069T3SAM Medium RangeE9AGM-88E527114750084750YesYesYes A070T3SAM Medium RangeE9AGM-88E60690000090000NoYesYes A071T3SAM Medium RangeE5GLSDB (GBU-39)80848000048000NoYesYes A072T3SAM Medium RangeE5GLSDB (GBU-39)821075000045000YesYesYes A073T3SAM Medium RangeE5GLSDB (GBU-39)1001067500067500NoYesYes A074T3SAM Medium RangeE5GLSDB (GBU-39)10414109500079500YesYesYes A075T3SAM Medium RangeE5GLSDB (GBU-39)1201290000090000NoYesYes A076T4SAM Short RangeE9AGM-88E40454000054000NoYesYes A077T4SAM Short RangeE9AGM-88E41582500052500YesYesYes A078T4SAM Short RangeE9AGM-88E50571250071250NoYesYes A079T4SAM Short RangeE9AGM-88E527114750084750YesYesYes A080T4SAM Short RangeE9AGM-88E60690000090000NoYesYes EXPLORING DECISION ADVANT AGES | 267 A081T4SAM Short RangeE5GLSDB (GBU-39)80848000048000NoYesYes A082T4SAM Short RangeE5GLSDB (GBU-39)821075000045000YesYesYes A083T4SAM Short RangeE5GLSDB (GBU-39)1001067500067500NoYesYes A084T4SAM Short RangeE5GLSDB

(GBU-39)10414109500079500YesYesYes A085T4SAM Short RangeE5GLSDB (GBU-39)1201290000090000NoYesYes A086T5SAM Long RangeE9AGM-88E40454000054000NoYesYes A087T5SAM Long RangeE9AGM-88E41582500052500YesYesYes A088T5SAM Long RangeE9AGM-88E50571250071250NoYesYes A089T5SAM Long RangeE9AGM-88E527114750084750YesYesYes A090T5SAM Long RangeE9AGM-88E60690000090000NoYesYes A091T6AirstripE8KEPD-3501010900000NoYesYes A092T6AirstripE7GBU-492020900000NoYesYes A093T6AirstripE6GBU-122020900000NoYesYes A094T6AirstripE11MQ-9 w GBU-122020900000NoYesYes A095T6AirstripE12MQ-9 w GBU-492020900000NoYesYes A096T7Artillery Unit Medium RangeE2Archer Bonus80848000048000NoYesYes A097T7Artillery Unit Medium RangeE2Archer Bonus841287000057000YesYesYes A098T7Artillery Unit Medium RangeE2Archer Bonus1001067500067500NoYesYes A099T7Artillery Unit Medium RangeE2Archer Bonus10616123000093000YesYesYes A100T7Artillery Unit Medium RangeE2Archer Bonus1201290000090000NoYesYes A101T7Artillery Unit Medium

RangeE7GBU-4910127000027000NoYesYes A102T7Artillery Unit Medium RangeE7GBU-4912396000066000YesYesYes A103T7Artillery Unit Medium RangeE7GBU-4920257000057000NoYesYes A104T7Artillery Unit Medium RangeE7GBU-49224129000099000YesYesYes A105T7Artillery Unit Medium RangeE7GBU-4930390000090000NoYesYes A106T7Artillery Unit Medium RangeE10TB2 MAM-L20240500040500NoYesYes A107T7Artillery Unit Medium RangeE10TB2 MAM-L22496000066000YesYesYes A108T7Artillery Unit Medium RangeE10TB2 MAM-L30364125064125NoYesYes A109T7Artillery Unit Medium RangeE10TB2 MAM-L325121875091875YesYesYes A110T7Artillery Unit Medium RangeE10TB2 MAM-L40490000090000NoYesYes A111T7Artillery Unit Medium RangeE12MQ-9 w GBU4910127000027000NoYesYes A112T7Artillery Unit Medium RangeE12MQ-9 w GBU4912396000066000YesYesYes A113T7Artillery Unit Medium RangeE12MQ-9 w GBU4920257000057000NoYesYes A114T7Artillery Unit Medium RangeE12MQ-9 w GBU49224129000099000YesYesYes A115T7Artillery Unit Medium RangeE12MQ-9 w GBU4930390000090000NoYesYes 268

| EXPLORING DECISION ADVANT AGES A116T7Artillery Unit Medium RangeE13MQ-9 w AGM-114 Hellfire II10127000027000NoYesYes A117T7Artillery Unit Medium RangeE13MQ-9 w AGM-114 Hellfire II12396000066000YesYesYes A118T7Artillery Unit Medium RangeE13MQ-9 w AGM-114 Hellfire II20257000057000NoYesYes A119T7Artillery Unit Medium RangeE13MQ-9 w AGM-114 Hellfire II224129000099000YesYesYes A120T7Artillery Unit Medium RangeE13MQ-9 w AGM-114 Hellfire II30390000090000NoYesYes A121T8Artillery Unit Short RangeE2Archer Bonus80848000048000NoYesYes A122T8Artillery Unit Short RangeE2Archer Bonus841287000057000YesYesYes A123T8Artillery Unit Short RangeE2Archer Bonus1001067500067500NoYesYes A124T8Artillery Unit Short RangeE2Archer Bonus10616123000093000YesYesYes A125T8Artillery Unit Short RangeE2Archer Bonus1201290000090000NoYesYes A126T8Artillery Unit Short RangeE13MQ-9 w AGM-114 Hellfire II10127000027000NoYesYes A127T8Artillery Unit Short RangeE13MQ-9 w AGM-114 Hellfire II12396000066000YesYesYes

A128T8Artillery Unit Short RangeE13MQ-9 w AGM-114 Hellfire II20257000057000NoYesYes A129T8Artillery Unit Short RangeE13MQ-9 w AGM-114 Hellfire II224129000099000YesYesYes A130T8Artillery Unit Short RangeE13MQ-9 w AGM-114 Hellfire II30390000090000NoYesYes A131T8Artillery Unit Short RangeE10TB2 MAM-L20240500040500NoYesYes A132T8Artillery Unit Short RangeE10TB2 MAM-L22496000066000YesYesYes A133T8Artillery Unit Short RangeE10TB2 MAM-L30364125064125NoYesYes A134T8Artillery Unit Short RangeE10TB2 MAM-L325121875091875YesYesYes A135T8Artillery Unit Short RangeE10TB2 MAM-L40490000090000NoYesYes A136T9Attack Helo SiteE8KEPD-35010142750042750NoYesYes A137T9Attack Helo SiteE8KEPD-350112100500070500YesYesYes A138T9Attack Helo SiteE8KEPD-3501231432500113250YesYesYes A139T9Attack Helo SiteE8KEPD-35020290000090000NoYesYes A140T9Attack Helo SiteE4ATACMS MGM 140B10142750042750NoYesYes A141T9Attack Helo SiteE4ATACMS MGM 140B112100500070500YesYesYes A142T9Attack Helo SiteE4ATACMS MGM

140B1231432500113250YesYesYes A143T9Attack Helo SiteE4ATACMS MGM 140B20290000090000NoYesYes A144T9Attack Helo SiteE7GBU-4940454000054000NoYesYes A145T9Attack Helo SiteE7GBU-4942696000066000YesYesYes A146T9Attack Helo SiteE7GBU-4950571250071250NoYesYes A147T9Attack Helo SiteE7GBU-49538129000099000YesYesYes EXPLORING DECISION ADVANT AGES | 269 A148T9Attack Helo SiteE7GBU-4960690000090000NoYesYes A149T9Attack Helo SiteE6GBU-1260643200043200NoYesYes A150T9Attack Helo SiteE6GBU-12641087000057000YesYesYes A151T9Attack Helo SiteE6GBU-1280864800064800NoYesYes A152T9Attack Helo SiteE6GBU-128513120300090300YesYesYes A153T9Attack Helo SiteE6GBU-121001090000090000NoYesYes A154T9Attack Helo SiteE11MQ-9 w GBU-1260643200043200NoYesYes A155T9Attack Helo SiteE11MQ-9 w GBU-12641087000057000YesYesYes A156T9Attack Helo SiteE11MQ-9 w GBU-1280864800064800NoYesYes A157T9Attack Helo SiteE11MQ-9 w GBU-128513120300090300YesYesYes A158T9Attack Helo SiteE11MQ-9 w GBU-121001090000090000NoYesYes A159T9Attack

Helo SiteE12MQ-9 w GBU-4940454000054000NoYesYes A160T9Attack Helo SiteE12MQ-9 w GBU-4942696000066000YesYesYes A161T9Attack Helo SiteE12MQ-9 w GBU-4950571250071250NoYesYes A162T9Attack Helo SiteE12MQ-9 w GBU-49538129000099000YesYesYes A163T9Attack Helo SiteE12MQ-9 w GBU-4960690000090000NoYesYes A164T9Attack Helo SiteE5GLSDB (GBU-39)40454000054000NoYesYes A165T9Attack Helo SiteE5GLSDB (GBU-39)42696000066000YesYesYes A166T9Attack Helo SiteE5GLSDB (GBU-39)50571250071250NoYesYes A167T9Attack Helo SiteE5GLSDB (GBU-39)538129000099000YesYesYes A168T9Attack Helo SiteE5GLSDB (GBU-39)60690000090000NoYesYes A169T9Attack Helo SiteE13MQ-9 w AGM-114 Hellfire II40454000054000NoYesYes A170T9Attack Helo SiteE13MQ-9 w AGM-114 Hellfire II42696000066000YesYesYes A171T9Attack Helo SiteE13MQ-9 w AGM-114 Hellfire II50571250071250NoYesYes A172T9Attack Helo SiteE13MQ-9 w AGM-114 Hellfire II538129000099000YesYesYes A173T9Attack Helo SiteE13MQ-9 w AGM-114 Hellfire II60690000090000NoYesYes A174T10Tank UnitE2Archer

Bonus80848000048000NoYesYes A175T10Tank UnitE2Archer Bonus841287000057000YesYesYes A176T10Tank UnitE2Archer Bonus1001067500067500NoYesYes A177T10Tank UnitE2Archer Bonus10616123000093000YesYesYes A178T10Tank UnitE2Archer Bonus1201290000090000NoYesYes A179T10Tank UnitE10TB2 MAM-L20240500040500NoYesYes A180T10Tank UnitE10TB2 MAM-L22496000066000YesYesYes A181T10Tank UnitE10TB2 MAM-L30364125064125NoYesYes A182T10Tank UnitE10TB2 MAM-L325121875091875YesYesYes A183T10Tank UnitE10TB2 MAM-L40490000090000NoYesYes A184T10Tank UnitE13MQ-9 w AGM-114 Hellfire II10127000027000NoYesYes A185T10Tank UnitE13MQ-9 w AGM-114 Hellfire II12396000066000YesYesYes A186T10Tank UnitE13MQ-9 w AGM-114 Hellfire II20257000057000NoYesYes A187T10Tank UnitE13MQ-9 w AGM-114 Hellfire II224129000099000YesYesYes 270 | EXPLORING DECISION ADVANT AGES A188T10Tank UnitE13MQ-9 w AGM-114 Hellfire II30390000090000NoYesYes A189T11SS Missile Long RangeE8KEPD-35030348600048600NoYesYes A190T11SS Missile Long

RangeE8KEPD-35032596000066000YesYesYes A191T11SS Missile Long RangeE8KEPD-35040468400068400NoYesYes A192T11SS Missile Long RangeE8KEPD-350426117600087600YesYesYes A193T11SS Missile Long RangeE8KEPD-35050590000090000NoYesYes A194T11SS Missile Long RangeE9AGM-88E30348600048600NoYesYes A195T11SS Missile Long RangeE9AGM-88E32596000066000YesYesYes A196T11SS Missile Long RangeE9AGM-88E40468400068400NoYesYes A197T11SS Missile Long RangeE9AGM-88E426117600087600YesYesYes A198T11SS Missile Long RangeE9AGM-88E50590000090000NoYesYes A199T11SS Missile Long RangeE4ATACMS MGM 140B30348600048600NoYesYes A200T11SS Missile Long RangeE4ATACMS MGM 140B32596000066000YesYesYes A201T11SS Missile Long RangeE4ATACMS MGM 140B40468400068400NoYesYes A202T11SS Missile Long RangeE4ATACMS MGM 140B426117600087600YesYesYes A203T11SS Missile Long RangeE4ATACMS MGM 140B50590000090000NoYesYes A204T11SS Missile Long RangeE11MQ-9 w GBU-1260643200043200NoYesYes A205T11SS Missile Long RangeE11MQ-9 w

GBU-12641087000057000YesYesYes A206T11SS Missile Long RangeE11MQ-9 w GBU-1280864800064800NoYesYes A207T11SS Missile Long RangeE11MQ-9 w GBU-128513120300090300YesYesYes A208T11SS Missile Long RangeE11MQ-9 w GBU-121001090000090000NoYesYes A209T11SS Missile Long RangeE12MQ-9 w GBU-4930348600048600NoYesYes A210T11SS Missile Long RangeE12MQ-9 w GBU-4932596000066000YesYesYes A211T11SS Missile Long RangeE12MQ-9 w GBU-4940468400068400NoYesYes A212T11SS Missile Long RangeE12MQ-9 w GBU-49426117600087600YesYesYes A213T11SS Missile Long RangeE12MQ-9 w GBU-4950590000090000NoYesYes A214T11SS Missile Long RangeE13MQ-9 w AGM-114 Hellfire II30348600048600NoYesYes A215T11SS Missile Long RangeE13MQ-9 w AGM-114 Hellfire II32596000066000YesYesYes A216T11SS Missile Long RangeE13MQ-9 w AGM-114 Hellfire II40468400068400NoYesYes A217T11SS Missile Long RangeE13MQ-9 w AGM-114 Hellfire II426117600087600YesYesYes A218T11SS Missile Long RangeE13MQ-9 w AGM-114 Hellfire II50590000090000NoYesYes A219T12Logistics depot

LargeE8KEPD-35010142750042750NoYesYes A220T12Logistics depot LargeE8KEPD-350112100500070500YesYesYes A221T12Logistics depot LargeE8KEPD-3501231432500113250YesYesYes A222T12Logistics depot LargeE8KEPD-35020290000090000NoYesYes A223T12Logistics depot LargeE4ATACMS MGM 140B10142750042750NoYesYes EXPLORING DECISION ADVANT AGES | 271 A224T12Logistics depot LargeE4ATACMS MGM 140B112100500070500YesYesYes A225T12Logistics depot LargeE4ATACMS MGM 140B1231432500113250YesYesYes A226T12Logistics depot LargeE4ATACMS MGM 140B20290000090000NoYesYes A227T12Logistics depot LargeE7GBU-4920240500040500NoYesYes A228T12Logistics depot LargeE7GBU-4922496000066000YesYesYes A229T12Logistics depot LargeE7GBU-4930364125064125NoYesYes A230T12Logistics depot LargeE7GBU-49325121875091875YesYesYes A231T12Logistics depot LargeE7GBU-4940490000090000NoYesYes A232T12Logistics depot LargeE6GBU-1240436000036000NoYesYes A233T12Logistics depot LargeE6GBU-1244887000057000YesYesYes A234T12Logistics depot

LargeE6GBU-1260660750060750NoYesYes A235T12Logistics depot LargeE6GBU-126511126375096375YesYesYes A236T12Logistics depot LargeE6GBU-1280890000090000NoYesYes A237T12Logistics depot LargeE11MQ-9 w GBU-1240436000036000NoYesYes A238T12Logistics depot LargeE11MQ-9 w GBU-1244887000057000YesYesYes A239T12Logistics depot LargeE11MQ-9 w GBU-1260660750060750NoYesYes A240T12Logistics depot LargeE11MQ-9 w GBU-126511126375096375YesYesYes A241T12Logistics depot LargeE11MQ-9 w GBU-1280890000090000NoYesYes A242T12Logistics depot LargeE12MQ-9 w GBU-4920240500040500NoYesYes A243T12Logistics depot LargeE12MQ-9 w GBU-4922496000066000YesYesYes A244T12Logistics depot LargeE12MQ-9 w GBU-4930364125064125NoYesYes A245T12Logistics depot LargeE12MQ-9 w GBU-49325121875091875YesYesYes A246T12Logistics depot LargeE12MQ-9 w GBU-4940490000090000NoYesYes A247T13Logistics depot MediumE7GBU-4910142750042750NoYesYes A248T13Logistics depot MediumE7GBU-49112100500070500YesYesYes A249T13Logistics depot

MediumE7GBU-491231432500113250YesYesYes A250T13Logistics depot MediumE7GBU-4920290000090000NoYesYes A251T13Logistics depot MediumE6GBU-1220240500040500NoYesYes A252T13Logistics depot MediumE6GBU-1222496000066000YesYesYes A253T13Logistics depot MediumE6GBU-1230364125064125NoYesYes A254T13Logistics depot MediumE6GBU-12325121875091875YesYesYes A255T13Logistics depot MediumE6GBU-1240490000090000NoYesYes A256T13Logistics depot MediumE11MQ-9 w GBU-1220240500040500NoYesYes A257T13Logistics depot MediumE11MQ-9 w GBU-1222496000066000YesYesYes A258T13Logistics depot MediumE11MQ-9 w GBU-1230364125064125NoYesYes A259T13Logistics depot MediumE11MQ-9 w GBU-12325121875091875YesYesYes A260T13Logistics depot MediumE11MQ-9 w GBU-1240490000090000NoYesYes A261T13Logistics depot MediumE12MQ-9 w GBU-4910142750042750NoYesYes A262T13Logistics depot MediumE12MQ-9 w GBU-49112100500070500YesYesYes A263T13Logistics depot MediumE12MQ-9 w GBU491231432500113250YesYesYes 272 | EXPLORING DECISION ADVANTAGES

A264T13Logistics depot MediumE12MQ-9 w GBU-4920290000090000NoYesYes A265T13Logistics depot MediumE3Archer Excalibur20240500040500NoYesYes A266T13Logistics depot MediumE3Archer Excalibur22496000066000YesYesYes A267T13Logistics depot MediumE3Archer Excalibur30364125064125NoYesYes A268T13Logistics depot MediumE3Archer Excalibur325121875091875YesYesYes A269T13Logistics depot MediumE3Archer Excalibur40490000090000NoYesYes A270T13Logistics depot MediumE10TB2 MAM-L20240500040500NoYesYes A271T13Logistics depot MediumE10TB2 MAM-L22496000066000YesYesYes A272T13Logistics depot MediumE10TB2 MAM-L30364125064125NoYesYes A273T13Logistics depot MediumE10TB2 MAM-L325121875091875YesYesYes A274T13Logistics depot MediumE10TB2 MAM-L40490000090000NoYesYes A275T13Logistics depot MediumE5GLSDB (GBU-39)10127000027000NoYesYes A276T13Logistics depot MediumE5GLSDB (GBU-39)12396000066000YesYesYes A277T13Logistics depot MediumE5GLSDB (GBU-39)20257000057000NoYesYes A278T13Logistics depot MediumE5GLSDB

(GBU-39)224129000099000YesYesYes A279T13Logistics depot MediumE5GLSDB (GBU-39)30390000090000NoYesYes A280T14Bridge LargeE8KEPD-35010127000027000NoYesYes A281T14Bridge LargeE8KEPD-35012396000066000YesYesYes A282T14Bridge LargeE8KEPD-35020257000057000NoYesYes A283T14Bridge LargeE8KEPD-350224129000099000YesYesYes A284T14Bridge LargeE8KEPD-35030390000090000NoYesYes A285T14Bridge LargeE7GBU-4930348600048600NoYesYes A286T14Bridge LargeE7GBU-4932596000066000YesYesYes A287T14Bridge LargeE7GBU-4940468400068400NoYesYes A288T14Bridge LargeE7GBU-49426117600087600YesYesYes A289T14Bridge LargeE7GBU-4950590000090000NoYesYes A290T14Bridge LargeE6GBU-1230348600048600NoYesYes A291T14Bridge LargeE6GBU-1232596000066000YesYesYes A292T14Bridge LargeE6GBU-1240468400068400NoYesYes A293T14Bridge LargeE6GBU-12426117600087600YesYesYes A294T14Bridge LargeE6GBU-1250590000090000NoYesYes A295T14Bridge LargeE11MQ-9 w GBU-1230348600048600NoYesYes A296T14Bridge LargeE11MQ-9 w GBU-1232596000066000YesYesYes A297T14Bridge

LargeE11MQ-9 w GBU-1240468400068400NoYesYes A298T14Bridge LargeE11MQ-9 w GBU-12426117600087600YesYesYes A299T14Bridge LargeE11MQ-9 w GBU-1250590000090000NoYesYes A300T14Bridge LargeE12MQ-9 w GBU-4930348600048600NoYesYes A301T14Bridge LargeE12MQ-9 w GBU-4932596000066000YesYesYes A302T14Bridge LargeE12MQ-9 w GBU-4940468400068400NoYesYes A303T14Bridge LargeE12MQ-9 w GBU-49426117600087600YesYesYes A304T14Bridge LargeE12MQ-9 w GBU-4950590000090000NoYesYes A305T15Bridge MediumE8KEPD-35010142750042750NoYesYes EXPLORING DECISION ADVANT AGES | 273 A306T15Bridge MediumE8KEPD-350112100500070500YesYesYes A307T15Bridge MediumE8KEPD-3501231432500113250YesYesYes A308T15Bridge MediumE8KEPD-35020290000090000NoYesYes A309T15Bridge MediumE7GBU-4910127000027000NoYesYes A310T15Bridge MediumE7GBU-4912396000066000YesYesYes A311T15Bridge MediumE7GBU-4920257000057000NoYesYes A312T15Bridge MediumE7GBU-49224129000099000YesYesYes A313T15Bridge MediumE7GBU-4930390000090000NoYesYes A314T15Bridge

MediumE6GBU-1210127000027000NoYesYes A315T15Bridge MediumE6GBU-1212396000066000YesYesYes A316T15Bridge MediumE6GBU-1220257000057000NoYesYes A317T15Bridge MediumE6GBU-12224129000099000YesYesYes A318T15Bridge MediumE6GBU-1230390000090000NoYesYes A319T15Bridge MediumE11MQ-9 w GBU-1210127000027000NoYesYes A320T15Bridge MediumE11MQ-9 w GBU-1212396000066000YesYesYes A321T15Bridge MediumE11MQ-9 w GBU-1220257000057000NoYesYes A322T15Bridge MediumE11MQ-9 w GBU-12224129000099000YesYesYes A323T15Bridge MediumE11MQ-9 w GBU-1230390000090000NoYesYes A324T15Bridge MediumE12MQ-9 w GBU-4910127000027000NoYesYes A325T15Bridge MediumE12MQ-9 w GBU-4912396000066000YesYesYes A326T15Bridge MediumE12MQ-9 w GBU-4920257000057000NoYesYes A327T15Bridge MediumE12MQ-9 w GBU-49224129000099000YesYesYes A328T15Bridge MediumE12MQ-9 w GBU-4930390000090000NoYesYes A329T16Bridge SmallE7GBU-491010900000NoYesYes A330T16Bridge SmallE6GBU-121010900000NoYesYes A331T16Bridge SmallE11MQ-9 w GBU-121010900000NoYesYes A332T16Bridge

SmallE12MQ-9 w GBU-491010900000NoYesYes // solution (multi-objective optimal) with objective 5 Total number of Prio 1 targets killed is 5 Total number of Prio 2 targets killed is 4 Total number of targets killed is 16 Total cost in USD is 7655000 Target T1 is assigned to effector E5 with attack option A033 Target T2 is assigned to effector E5 with attack option A055 Target T3 is assigned to effector E5 with attack option A075 Target T4 is assigned to effector E5 with attack option A085 Target T5 is assigned to effector E9 with attack option A090 Target T6 is assigned to effector E6 with attack option A093 Target T7 is assigned to effector E10 with attack option A107 Target T8 is assigned to effector E10 with attack option A132 274 | EXPLORING DECISION ADVANTAGES Target T9 is assigned to effector E6 with attack option A153 Target T10 is assigned to effector E10 with attack option A183 Target T11 is assigned to effector E12 with attack option A213 Target T12 is assigned to effector

E7 with attack option A231 Target T13 is assigned to effector E12 with attack option A264 Target T14 is assigned to effector E6 with attack option A294 Target T15 is assigned to effector E6 with attack option A318 Target T16 is assigned to effector E6 with attack option A330 Kill of target T1 is 1 Realized kill probability of target T1 is 0.9 Kill of target T2 is 1 Realized kill probability of target T2 is 0.9 Kill of target T3 is 1 Realized kill probability of target T3 is 0.9 Kill of target T4 is 1 Realized kill probability of target T4 is 0.9 Kill of target T5 is 1 Realized kill probability of target T5 is 0.9 Kill of target T6 is 1 Realized kill probability of target T6 is 0.9 Kill of target T7 is 1 Realized kill probability of target T7 is 0.96 Kill of target T8 is 1 Realized kill probability of target T8 is 0.96 Kill of target T9 is 1 Realized kill probability of target T9 is 0.9 Kill of target T10 is 1 Realized kill probability of target T10 is 0.9 Kill of target T11 is 1

Realized kill probability of target T11 is 0.9 Kill of target T12 is 1 Realized kill probability of target T12 is 0.9 Kill of target T13 is 1 Realized kill probability of target T13 is 0.9 Kill of target T14 is 1 Realized kill probability of target T14 is 0.9 Kill of target T15 is 1 Realized kill probability of target T15 is 0.9 Kill of target T16 is 1 Realized kill probability of target T16 is 0.9 EXPLORING DECISION ADVANT AGES | 275 Total units of weapons E1 used is 0 Total units of weapons E2 used is 0 Total units of weapons E3 used is 0 Total units of weapons E4 used is 0 Total units of weapons E5 used is 30 Total units of weapons E6 used is 21 Total units of weapons E7 used is 4 Total units of weapons E8 used is 0 Total units of weapons E9 used is 6 Total units of weapons E10 used is 12 Total units of weapons E11 used is 0 Total units of weapons E12 used is 7 Total units of weapons E13 used is 0 276 | EXPLORING DECISION ADVANT AGES Appendix C – Example of script Setup

This is a fictive example of a dynamic targeting decision by the J3 Director (J3Dir) (who has a delegated target engagement authority by the joint force commander). There are four other persons involved in this script, and a fifth person mentioned but not present: A Targeteer from the joint fires element (JFE); an intelligence analyst (J2) from the J2 Intelligence directorate; a watch-keeper/duty-officer (DO); and a legal advisor (LEGAD) from special staff. The script deliberately contains relevant abbreviations, tactical terms etcetera, but also reflects the participants knowledge of the conversation being recorded. Script J3Dir: Ok, attention, listen up. An update on the evolving situation in vicinity of Haradshere (pointing at the big screen). DO, give us the latest DO: Sir, we are currently monitoring five objects – two of those are PID as Himars units, one as a potential MSAM site, and two fast-moving tank companies heading southeast as seen here (pointing at screen). J3Dir to

J2: Any updates on the MSAM? J2: No, Sir! J3Dir: What do we make of all this? J2: Well, if I maythe two Himars units are assessed to be there in support of the adversary’s advancement to the southeastwe assess that the two tank companies will try to make it over the two bridges hereand here and then secure a bridgehead until released. J3Dir: For the recordthese are bridges over Piteälven north and northwest of Älvsbyn. J2: Yes, that’s correct, Sir. J3Dir: Ok, got youso what is the estimated time before those tank companies reaches the two bridges? J2: Give me two minutes to find out Sir J3Dir: Is it possible for us to take out the bridges? JFE: Certainly, we have the capabilitybut I don’t know the status and specific whereabouts of our effectors in this particular area. Of course, we could always use two of the JAS-39 that are on stand-by, however they can be wheels-up in about 15minutes EXPLORING DECISION ADVANT AGES | 277 from Uppsala, but then it is the flight time to

get there, and so onHowever, we need to check if the bridges are on the JTL and if there are any restrictions! J3Dir: Get me LEGAD over here asap! J2: We estimate that the tank companies will reach the bridges within 20-30 min given their current speed J3Dir: Ok. Any update on the suspected MSAM? I bet it will push forward to protect the bridges once the tanks have made it over the river. JFE: Sir, both bridges are on the JTL. No restrictions J3Dir: LEGAD? LEGAD: Nothing to add from a legal perspective. However, destroying both these bridges will have a second order effect impacting the civilian population, especially that one (pointing at the screen), which is the larger of the two. J3Dir: You mean the one closest to Älvsbyn? LEGAD: Yes, I do. J3Dir: Has anyone seen POLAD? J3Dir: Okso given the circumstances and J2 assessment we need to be swift here. We have two critical target, potentially time-sensitive in the two tank companies. However, if we are able to destroy the two bridges,

those unit cannot make it over, which gives us time to engage them if needed. The MSAM, if it is an MSAM, pose a credible threat to any fighter aircraft going after the two HIMARS. OPTIMIZER! What options do we have if we try to DESTROY the two bridges, DELAY the two tank companies, and DISRUPT the HIMARS to ensure they would not counterattack any of our own ground forces used for targeting? By the way, DO, remind me to inform POLAD and InfoOps! Also, J2, make sure to have resources tasked to perform BDAs, especially the bridges. I need to know that they have been destroyed! J2: Already on it, Sir. Will confirm when any available collection asset is tasked! (OPTIMIZATION MODEL WORKING WITH THE INPUT (a second or two), and then the GenAI agent presents two options on the screen: Option A and B. Both meet the objective functions and finds a global optima.) J3Dir: Given the available options, Option A reduces the risk to own forces, because it uses long-range ground forces (GLSDB) and an

MQ-9 to DISRUPT the HIMARS. OPTIMIZER! Summarize the key concerns, and add what I just said about the risk, and then transform this into my targeting decision and print it so I can sign off. Any objections? Ok, no objections. End of script // 278 | EXPLORING DECISION ADVANTAGES References Abdollahi, Abolfazl, Biswajeet Pradhan, and Abdullah M. Alamri 2022 “An Ensemble Architecture of Deep Convolutional Segnet and Unet Networks for Building Semantic Segmentation from High-Resolution Aerial Images.” Geocarto International 37 (12): 3355–70 https://doi.org/101080/1010604920201856199 Ahuja, Ravindra K., Arvind Kumar, Krishna C Jha, and James B Orlin 2007 “Exact and Heuristic Algorithms for the Weapon-Target Assignment Problem.” Operations Research 55 (6): 1136–46 https://doi.org/101287/opre10700440 “AI Og Autonome Systemer.” 2024 Norwegian Air Power Journal 1 (1): 1–86 Al-Ababneh, Mukhles M. 2020 “Linking Ontology, Epistemology and Research Methodology.” Science

& Philosophy 8 (1): 75–91 https://doi.org/1023756/spv8i1500 Alberts, David S, and Richard E Hayes. 2003 Code of Best Practice: Experimentation. Information Age Vol 10 wwwdodccrporg Allen, John R., and Amir Husain 2017 “On Hyperwar” US Naval Institute 143/7 (July). https://www.usniorg/magazines/proceedings/2017/july/hyperwar Alowais, Shuroug A, Sahar S Alghamdi, Nada Alsuhebany, Tariq Alqahtani, Abdulrahman I Alshaya, Sumaya N Almohareb, Atheer Aldairem, et al. 2023. “Revolutionizing Healthcare: The Role of Artificial Intelligence in Clinical Practice.” BMC Medical Education 23 (1): 689 https://doi.org/101186/s12909-023-04698-z Andersen, Alexandre Colaers, Konstantin Pavlikov, and Túlio A.M Toffolo 2022. “Weapon-Target Assignment Problem: Exact and Approximate Solution Algorithms.” Annals of Operations Research 312 (2): 581–606 https://doi.org/101007/s10479-022-04525-6 Asdal, Kristin, Brita Brenna, and Ingunn Moser. 2007 “Technoscience The Politics of

Interventions.” wwwunipubno Asimov, Isaac. 2013 I, Robot Harper Voyager Ayoub, Kareem, and Kenneth Payne. 2016 “Strategy in the Age of Artificial Intelligence.” Journal of Strategic Studies 39 (5–6): 793–819 https://doi.org/101080/0140239020151088838 Babul Hasan, Mohammad, and Yaindrila Barua. 2021 “Weapon Target Assignment.” In Concepts, Applications and Emerging Opportunities in Industrial Engineering. https://doiorg/105772/intechopen93665 Barseghyan, Hakob. 2018 “Redrafting the Ontology of Scientific Change” Scientonomy: Journal for the Science of Science 2: 13–38. https://doi.org/1033137/jsv2i031032 Ben-Ari, Eyal. 2014 “Reflexivity: Potentially ‘Dangerous Liaisons’” Bookitem In Routledge Handbook of Research Methods in Military Studies, 29–39. Routledge. https://doiorg/104324/9780203093801-13 Ben-Ari, Eyal, and Yagil Levy. 2014 “Getting Access to the Field: EXPLORING DECISION ADVANT AGES | 279 Insider/Outsider Perspectives.” Routledge Handbook

of Research Methods in Military Studies, 9–18. https://doiorg/104324/9780203093801-11 Berman, Maxim, Amal Rannen, and Triki Matthew. 2018 “Berman The LovaSz-Softmax Loss CVPR 2018 paper.Pdf” Cvpr, 4413– 21. http://openaccess.thecvfcom/content cvpr 2018/html/Berman The Lo vaSz-Softmax Loss CVPR 2018 paper.html Bertsekas, Dimitri P. 2008 “Auction Algorithms” Encyclopedia of Optimization, 128–32. https://doiorg/101007/978-0-387-74759-0 22 Biasutti, Pierre, Vincent Lepetit, Jean Francois Aujol, Mathieu Bredif, and Aurelie Bugeau. 2019 “LU-Net: An Efficient Network for 3D LiDAR Point Cloud Semantic Segmentation Based on End-to-End-Learned 3D Features and U-Net.” Proceedings - 2019 International Conference on Computer Vision Workshop, ICCVW 2019, 942–50. https://doi.org/101109/ICCVW201900123 Biddle, Stephen. 2004 Explaining Victory and Defeat in Modern Battle . 2010 Military Power: Explaining Victory and Defeat in Modern Battle Military Power: Explaining Victory and

Defeat in Modern Battle. https://doi.org/102307/20034231 Biddle, Tami Davis. 2019 “Air Power and Warfare: A Century of Theory and History.” Strategic Studies Institute and US Army War College Press Bird, Alexander. 2022 “Thomas Kuhn” Stanford Encyclopedia of Philosophy 2022. https://platostanfordedu/archives/spr2022/entries/thomaskuhn/ Bode, Ingvild, and Hendrik Huelss. 2021 “Autonomous Weapons Systems and Changing Norms in International Relations.” Review of International Studies 44: 393–413. https://doiorg/101017/S0260210517000614 . 2024 “Artificial Intelligence Technologies and Practical Normativity/Normality: Investigating Practices beyond the Public Space.” Open Research Europe 3. https://doi.org/1012688/openreseurope165362 Bode, Ingvild, Hendrik Huelss, Anna Nadibaidze, Guangyu Qiao-Franco, and Tom F.A Watts 2024 “Algorithmic Warfare: Taking Stock of a Research Programme.” Global Society 38 (1): 1–23 https://doi.org/101080/1360082620232263473 Bode,

Ingvild, and Tom Watts. 2021 “Meaning-Less Human Control” Bogdanowicz, Zbigniew R. 2009 “A New Efficient Algorithm for Optimal Assignment of Smart Weapons to Targets.” Computers and Mathematics with Applications 58 (10): 1965–69. https://doi.org/101016/jcamwa200907082 Borchert, Heiko, Christian Brandlhuber, Armin Brandstetter, and Gary S. Schaal. 2022 “Free Jazz on the Battlefield” Borchert, Heiko, Torben Schutz, and Joseph Verbovszky. 2024 “Master and Servant: Defense AI in Germany.” In The Very Long Game, edited by Heiko Borchert, Torben Schutz, and Joseph Verbovszky, 195–216. Cham, Switzerland: Springer Nature. https://doiorg/101007/978-3-03158649-1 280 | EXPLORING DECISION ADVANT AGES Boulanin, Vincent. 2019 “The Impact of Artificial Intelligence on Strategic Stability and Nuclear Risk - Euro-Atlantic Perspectives.” Sipri Vol I https://www.sipriorg/publications/2019/other-publications/impactartificial-intelligence-strategic-stability-and-nuclear-risk

Boulanin, Vincent, and Maaike Verbruggen. 2017 “MAPPING THE DEVELOPMENT OF AUTONOMY IN WEAPON SYSTEMS Vincent Boulanin and Maaike Verbruggen.” Boury-Brisset, Anne-Claire, and Jean Berger. 2020 “Benefits and Challenges of AI/ML in Support of Intelligence and Targeting in Hybrid Military Operations,” 1–12. Bousquet, Antoine. 2023 “The Persistent Appeal of Chaoplexic Warfare: Towards an Autonomous S(War)m Machine?” In Routledge Handbook of the Future of Warfare, 1st Editio, 11. Routledge Bovet Emanuel, Peter. 2024 “Making Chameleons: Techno-Shaped, SoftwareDefined Special Operations Assemblages” In Into the Void, edited by James D. Kiras and Martijn Kitzen, 394 London, England: Hurst & Co Ltd Box, G E P, and N R Draper. 1987 Empirical Model-Building and Response Surfaces. Wiley Series in Probability and Statistics - Applied Probability and Statistics Section Series. Wiley https://books.googlese/books?id=QO2dDRufJEAC Boyd, John R. 2018 A Discourse on Winning and

Losing Edited by Grant T Hammond. Maxwell AFB, Alabama: Air University Press . 2020 Patterns of Conflict Edited by Chet Richards and Chuck Spinney Defense in the national interest. dnipogoorg Boyd, John R. 1976 “V - Destruction and Creation” A Discourse on Winning and Losing, 1–9. Brehmer, Berndt. 2005 “The Dynamic OODA Loop: Amalgamating Boyd’s OODA Loop and the Cybernetic Approach to Command and Control.” 10th International Command and Control Research and Technology Symposium The Future of C2, 365–68. . 2010 “Command and Control as Design” Proceedings of the 15th International Command and Control Research and Technology Symposium. Brent, Keith, ed. 2005 Masters of Air Power 2005 RAAF History Conference 2nd ed. Canberra: Air Power Development Centre TCC-3, Department of Defence. Breton, Richard, and Robert Rousseau. 2005 “The Future of C 2 THE C-OODA : A COGNITIVE VERSION OF THE OODA LOOP TO REPRESENT C 2 ACTIVITIES.” In The Future of C 2 THE C-OODA : A

COGNITIVE VERSION OF THE OODA LOOP TO REPRESENT C 2 ACTIVITIES. Brister, Paul. 2021 “Making Time an Ally” In Wartime - Temporality and the Decline of Western Military Power, edited by Sten Rynning, Olivier Schmitt, and Amelie Theussen, 316. Washington, D C: Brookings Institution Press. Brodie, B. 1943 A Layman’s Guide to Naval Strategy Princeton University Press. https://booksgooglese/books?id=r5QsAAAAYAAJ Brose, Christian. 2020 The Kill Chain: Defending America in the Future of HighTech Warf New York, USA: Hachette Books EXPLORING DECISION ADVANT AGES | 281 Buchanan, Bruce G., and Reid G Smith 1987 “Fundamentals of Expert System” https://doi.org/101007/978-3-662-44497-9 3 Budning, Kevin, Alex Wilner, and Guillaume Cote. 2021 “Connecting the Dots on Canada’s Connected Battlespace.” International Journal 76 (1): 154– 62. https://doiorg/101177/0020702021992855 Cai, Hongxia, Yunqi Bian, and Lilan Liu. 2023 “Deep Reinforcement Learning for Solving Resource

Constrained Project Scheduling Problems with Resource Disruptions.” https://doiorg/101016/jrcim2023102628 Cave, Stephen, and Seán S ÓhÉigeartaigh. 2018 “An AI Race for Strategic Advantage: Rhetoric and Risks.” https://doi.org/101145/32787213278780 Chairman of the Joint Chiefs of Staff. 2021 “DOD Dictionary of Military and Associated Terms.” Joint Education and Doctrine Division, J-7, no January: 382. https://www.jcsmil/Portals/36/Documents/Doctrine/pubs/dictionary pdf. Chamola, Vinay, Vikas Hassija, A. Razia Sulthana, Debshishu Ghosh, Divyansh Dhingra, and Biplab Sikdar. 2023 “A Review of Trustworthy and Explainable Artificial Intelligence (XAI).” IEEE Access 11 (Ml): 1–23 https://doi.org/101109/ACCESS20233294569 Chang, Xuening, Jianmai Shi, Zhihao Luo, and Yao Liu. 2023a “Adaptive Large Neighborhood Search Algorithm for Multi-Stage Weapon Target Assignment Problem.” Computers and Industrial Engineering 181 (July) https://doi.org/101016/JCIE2023109303 . 2023b

“Adaptive Large Neighborhood Search Algorithm for MultiStage Weapon Target Assignment Problem” Computers and Industrial Engineering 181 (July). https://doiorg/101016/jcie2023109303 Chattoe, Edmund, Nicole J Saam, and Michael Möhring. 2000 “Sensitivity Analysis in the Social Sciences: Problems and Prospects BT - Tools and Techniques for Social Science Simulation.” In , edited by Ramzi Suleiman, Klaus G Troitzsch, and Nigel Gilbert, 243–73. Heidelberg: Physica-Verlag HD. Clark, Bryan. 2019 “Mosaic Warfare-Exploiting Artificial Intelligence and Autonomous Systems to Inplement Decision-Centric Operations” 102 (11): 51–55. Clark, Robert M. 2020 Intelligence Analysis - a Target-Centric Appraoch Sixth edit. Thousand Oaks: SAGE Publications Inc Cococcioni, Marco, Massimo Pappalardo, and Yaroslav D. Sergeyev 2018 “Lexicographic Multi-Objective Linear Programming Using Grossone Methodology: Theory and Algorithm.” Applied Mathematics and Computation 318: 298–311.

https://doiorg/101016/jamc201705058 Cohen, Louis, Manion Lawrence, and Keith Morrison. 2011 Research Methods in Education. Seventh Edition Research Methods in Education Cohen, Mika, Farzad Kamrani, Fredrik Bissmarck, and Peter Hammar. 2020 “Krigsspel Med AlphaZero,” no. December Coram, Robert. 2002 The Fighter Pilot Who Changed the Art of War New York: Back Bay Books/ Little, Brown and Company. 282 | EXPLORING DECISION ADVANT AGES Creveld, Martin van. 1985 Command in War Cambridge, Mass ; Oxford [England] : Harvard University Press. . 1989 Technology and War: From 2000 BC to the Present New York, NY, USA: The Free Press. . 2020 Seeing into the Future London, England: Reaktion Books Ltd Cronin, Audrey Kurth. 2019 “Power to the People: How Open Technological Innovation Is Arming Tomorrow’s Terrorists.” Oxford University Press (OUP).

https://newbooksnetworkcom/audrey-kurth-cronin-power-tothe-people-how-open-technological-innovation-is-arming-tomorrowsterrorists-oxford-up-2019/ Danzig, George B. 1963 “Linear Programming and Extentions; R-366-PR” Davies, Harry, Bethan Mckernan, and Dan Sabbagh. 2023 “‘The Gospel’: How Israel Uses AI to Select Bombing Targets in Gaza.” The Guardian, 2023 DCDC. 2020 “JCN 1/20 Multi-Domain Integration,” 1–90 https://dcdc.moduk/content/joint concepts/pubs/20201112JCN 1 20 MDI Securepdf Defense Advanced Research Projects Agency. 2018a “AI Next Campaign” 2018. https://wwwdarpamil/work-with-us/ai-next-campaign . 2018b “DARPA Tiles Together a Vision of Mosaic Warfare” https://www.darpamil/work-with-us/darpa-tiles-together-a-vision-ofmosiac-warfare DeLanda, Manuel. 1991 War in the Age of Intelligent Machines Zone Books Delft, TU. 2023 “Responsible AI in the Military Domain” Reaim 2023 https://www.governmentnl/ministries/ministry-of-foreignaffairs/events/reaim

Delgado-Rodríguez, Miguel, and Javier Llorca. 2004 “Bias” Journal of Epidemiology and Community Health 58 (8): 635–41. https://doi.org/101136/jech2003008466 Department of Defense. 2022 “Summary of the Joint All-Domain Command & Control (Jadc2) Strategy,” no. March https://media.defensegov/2022/Mar/17/2002958406/-1/1/1/SUMMARY-OF-THE-JOINT-ALL-DOMAIN-COMMAND-ANDCONTROL-STRATEGYPDF Deptula, David A. 2001 “Effects-Based Operations: CHANGE IN THE NATURE OF WARFARE.” wwwAEForg Derian, James Der. 1990 “The ( S ) Pace of International Relations : Simulation , Surveillance , and Speed.” International Studies 34 (3): 295–310 Derian, James Der, and Stuart Rollo. 2023 “Quantum Warfare” Routledge Handbook of the Future of Warfare, no. March: 319–29 https://doi.org/104324/9781003299011-34 Douhet, Giulio. 2019 The Command of the Air (D Ferrari, Trad) Alabama: Air University Press. Druckman, James N. 2022 Experimental Thinking Cambridge University Press

https://doi.org/101017/9781108991353 Du, Hung, Srikanth Thudumu, Rajesh Vasa, and Kon Mouzakis. 2024 “A Survey on Context-Aware Multi-Agent Systems: Techniques, Challenges and Future Directions.” http://arxivorg/abs/240201968 EXPLORING DECISION ADVANT AGES | 283 Ducheine, Paul A.L, Michael N Schmitt, and Frans PB Osinga, eds 2015 Targeting: The Challenges of Modern Warfare. Book Targeting: The Challenges of Modern Warfare. 1st ed 20 The Hague: TMC Asser Press https://doi.org/101007/978-94-6265-072-5 Dunnmon, Jared, Bryce Goodman, Peter Kirechu, Carol Smith, and Alexandrea Van Deusen. 2021 Responsible AI Guidelines in Practice US Defense Innovation Unit. Eckel, Mike. 2020 “Drone Wars: In Nagorno-Karabakh, The Future Of Warfare Is Now.” 2020 https://wwwrferlorg/a/drone-wars-in-nagornokarabakh-the-future-of-warfare-is-now/30885007html Ekelhof, Merel. 2018 “Lifting the Fog of Targeting” Naval War College Review 71 (3): 61–95. . 2019 “Moving Beyond Semantics on

Autonomous Weapons: Meaningful Human Control in Operation.” Global Policy 10 (3): 343–48 https://doi.org/101111/1758-589912665 Engelbart, Doug. 1962 “Augmenting Human Intellect: A Conceptual Framework.” Eric Schmidt, Robert Work, Safra Catz, Eric Horovitz, Steve Chien, Andrew Jassy, Mignon Clyburn, et al. 2021 “Final Report of the National Security Commission on Artificial Intelligence.” National Security Commission on Artificial Intelligence. https://wwwnscaigov/2021-final-report/ Ernst, Dieter. 2020 “Competing in Artificial Intelligence Chips: China’s Challenge amid Technology War.” EURO. 2024 “33rd European Conference on Operational Research” In , 701 Evera, Stephen Van. 1997 Guide to Methods for Students of Political Science New York: Cornell University Press. Fan, Lili, Jiabin Yuan, Xuewei Niu, Keke Zha, and Weiqi Ma. 2023 “RockSeg: A Novel Semantic Segmentation Network Based on a Hybrid Framework Combining a Convolutional Neural Network and Transformer for

Deep Space Rock Images.” Remote Sensing 15 (16) https://doi.org/103390/rs15163935 Finlan, Alastair. 2008 Special Forces, Strategy and the War on Terror New York, USA: Routledge. . 2021 “The Shape of Warfare to Come: A Swedish Perspective 20202045” Article Defense & Security Analysis 37 (4): 472–91 https://doi.org/101080/1475179820211995976 Finlan, Alastair, Anna Danielsson, and Stefan Lundqvist. 2021 “Critically Engaging the Concept of Joint Operations: Origins, Reflexivity and the Case of Sweden.” Defence Studies 00 (00): 1–19 https://doi.org/101080/1470243620211932476 FOI. 2020 “Military Technology 2045 - A Basis for the Armed Forces Long Term Planning Process.” FOI-R--4985--SE, ISSN 1650-1942 Frankish, Keith, and William M. Ramsey, eds 2014 The Cambridge Handbook of Artificial Intelligence. Cambridge: Cambridge University Press Freedman, Lawrence. 2017 The Future of War : A History Allen Lane UK : Allen Lane. Friedrichs, Jrg, and Friedrich Kratochwil. 2009

“On Acting and Knowing: How 284 | EXPLORING DECISION ADVANT AGES Pragmatism Can Advance International Relations Research and Methodology.” International Organization 63 (4): 701–31 https://doi.org/101017/S0020818309990142 George, Alexander L., and Andrew Bennett 2005 Case Studies and Theory Development in the Social Sciences. MIT Press Journals Gevarter, William B. 1982 “An Overview of Expert Systems” Gill, Amandeep Singh. 2019 “Artificial Intelligence and International Security: The Long View.” Article Ethics and International Affairs 33 (2): 169–79 https://doi.org/101017/S0892679419000145 . 2020 “ROUNDTABLE: AI AND THE FUTURE OF GLOBAL AFFAIRS Artificial Intelligence and International Security: The Long View.” https://doi.org/101017/S0892679419000145 Glock, John R. 2012 “The Evolution of Air Force Targeting” Air and Space Power Journal 26 (6): 146–74. Golafshani, Nahid. 2003 “Understanding Reliability and Validity in Qualitative Research.” The

Qualitative Report 8 (4): 597 https://doi.org/1046743/2160-3715/20031870 Goodfellow, Ian, Yoshua Bengio, and Aaron Courville. 2016 Deep Learning MIT Press. Goyal, Anirudh, and Yoshua Bengio. 2020 “Inductive Biases for Deep Learning of Higher-Level Cognition.” Article Grant, Tim. 2023 “Detect, Deny, Degrade, Disrupt, Destroy, Deceive: Which Is the Greatest in OCO-.” European Conference on Information Warfare and Security, ECCWS 2023-June: 214–22. https://doi.org/1034190/eccws2211089 Gray, Colin S. 1999 Modern Strategy Oxford, England: Oxford University Press . 2009 “Understanding Airpower” Air Force Research Institute Papers Alabama. http://wwwjstororg/stable/26268353 Green, Jon, and Mark H. I I White 2023 “Machine Learning for Experiments in the Social Sciences.” Elements in Experimental Political Science Cambridge: Cambridge University Press. https://doiorg/DOI: 10.1017/9781009168236 Griffith, Samuel B. 1963 Sun Tzu: The Art of War Oxford, England: Oxford University

Press. Guillemin, Marilys, and Lynn Gillam. 2004 “Ethics, Reflexivity, and ‘Ethically Important Moments’ in Research.” Qualitative Inquiry 10 (2): 261–80 https://doi.org/101177/1077800403262360 Hamm, Manfred R. 1988 “The Airland Battle Doctrine: NATO Strategy and Arms Control in Europe.” Comparative Strategy 7 (3): 183–211 https://doi.org/101080/01495938808402742 Hammes, T.X 2021 “The Kill Chain: Defending America in the Future of HighTech Warf” National Defense University Press Vol PRISM Hammond, Grant. 2001 The Mind of War: John Boyd and American Security Smithsonian Institution. https://books.googlecom/books?hl=en&lr=&id=5Z9fBgAAQBAJ&oi=fnd &pg=PT5&ots=0u8GvBHs9O&sig=WR1msXIzoOWtT0nkO0UMHBDO0vU #v=onepage&q&f=false. EXPLORING DECISION ADVANT AGES | 285 He, Kaiming, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. 2015 “Deep Residual Learning for Image Recognition.” Proceedings of the IEEE Computer Society Conference on Computer

Vision and Pattern Recognition 2016December (December): 770–78. https://doiorg/101109/CVPR201690 Hendrickson, Katherine, Prashant Ganesh, Kyle Volle, Paul Buzaud, Kevin Brink, and Matthew Hale. 2023 “Decentralized Weapon–Target Assignment Under Asynchronous Communications.” Journal of Guidance, Control, and Dynamics 46 (2): 312–24. https://doi.org/102514/1G006532 Hill, Raymond R., and J O Miller 2017 “A HISTORY OF UNITED STATES MILITARY SIMULATION.” In Proceedings of the 2017 Winter Simulation Conference, edited by Eds. W K V Chan, A D’Ambrogio, G Zacharewicz, N. Mustafee, G Wainer and E Page, eds, 1–19 IEEE https://doi.org/105555/32421813242204 Hillier, Frederick S., and Gerald J Lieberman 2021 Introduction to Operations Research. Eleventh e New York, NY, USA: McGraw-Hill Education, Hinton, Geoffrey, Simon Osindero, and Yee-Whye Teh. 2006 “A Fast Learning Algorithm for Deep Belief Nets.” International Journal of Environmental Science and Development, no. Neural

Computation 18: 1527–54 https://doi.org/107763/ijesd2010v167 Hoadley, Daniel S., and Nathan L Lucas 2018 “Artificial Intelligence and National Security.” Congressional Research Service wwwcrsgovR45178 Hocaoğlu, Mehmet Fatih. 2019 “Weapon Target Assignment Optimization for Land Based Multi-Air Defense Systems: A Goal Programming Approach.” Computers and Industrial Engineering 128 (February): 681–89. https://doi.org/101016/jcie201901015 Hoehn, John R. 2021 “Joint All-Domain Command and Control : Background and Issues for Congress.” Congressional Research Service Vol R46725 Høiback, Harald. 2011 “What Is Doctrine?” Journal of Strategic Studies 34 (6): 879–900. https://doiorg/101080/014023902011561104 Holland Michel, Arthur. 2023 “War Machines” MIT Technology Review, 2023 https://www.technologyreviewcom/2023/08/16/1077386/warmachines/ Horn, Bernd, and Hans Ilis-Alm, eds. 2024 FORCE MULTIPLIER: UTILIZATION OF SOF FROM A SMALL STATE PERSPECTIVE. Ottawa, Ontario:

17 Wing Winnipeg Publishing Office. Horowitz, Michael C. 2016 “The Ethics & Morality of Robotic Warfare: Assessing the Debate over Autonomous Weapons.” Daedalus 145 (4) https://doi.org/101162/DAED a 00409 Horowitz, Michael C, Gregory C Allen, Elsa B Kania, and Paul Scharre. 2018 “Strategic Competition in an Era of Artificial Intelligence.” Husain, A, J R Allen, and R O Work. 2024 Hyperwar: Conflict and Competition in the AI Century. AM Press https://books.googlecom/books?id=qH510AEACAAJ IBM Corporation. 2013 “IBM ILOG CPLEX Optimization Studio CPLEX User’s Manual, Version 12 Release 6,” 1–468. Jackson, Ilya, Dmitry Ivanov, Alexandre Dolgui, and Jafar Namdar. 2024 286 | EXPLORING DECISION ADVANT AGES “Generative Artificial Intelligence in Supply Chain and Operations Management: A Capability-Based Framework for Analysis and Implementation.” International Journal of Production Research, 1–26 https://doi.org/101080/0020754320242309309 Jacobson, Ivar, Ian

Spence, and Brian Kerr. 2016 “Use-Case 20: The Hub of Software Development.” Queue 14 (1): 94–123 https://doi.org/101145/28984422912151 Jensen, Benjamin, Scott Cuomo, and Chris Whyte. 2018 “Wargaming with Athena: How to Make Militaries Smarter, Faster, and More Efficient with Artificial Intelligence.” War on the Rocks 2018 https://warontherocks.com/2018/06/wargaming-with-athena-how-tomake-militaries-smarter-faster-and-more-efficient-with-artificialintelligence/ Jensen, Benjamin M, Christopher Whyte, and Scott Cuomo. 2019 “Algorithms at War: The Promise, Peril, and Limits of Artificial Intelligence.” International Studies Review, June. https://doiorg/101093/isr/viz025 Jiang, Yuchen, Xiang Li, Hao Luo, Shen Yin, and Okyay Kaynak. 2022 “Quo Vadis Artificial Intelligence?” Discover Artificial Intelligence 2 (1). https://doi.org/101007/s44163-022-00022-8 Johnson, Bonnie, John M Green, Gregory Burns, Todd Collier, and Richard Cornish. 2023 “Mapping Artificial Intelligence

to the Naval Tactical Kill Chain.” NAVAL ENGINEERS JOURNAL No 135-1 (March): 155–67 Johnson, James. 2019a “Artificial Intelligence & Future Warfare: Implications for International Security.” Defense and Security Analysis 35 (2): 147–69 https://doi.org/101080/1475179820191600800 . 2019b “The End of Military-Techno Pax Americana? Washington’s Strategic Responses to Chinese AI-Enabled Military Technology.” Pacific Review. https://doiorg/101080/0951274820191676299 . 2019c “The AI-Cyber Nexus: Implications for Military Escalation, Deterrence and Strategic Stability.” Journal of Cyber Policy 4 (3): 442–60 https://doi.org/101080/2373887120191701693 . 2020a “Delegating Strategic Decision-Making to Machines: Dr Strangelove Redux?” Journal of Strategic Studies. https://doi.org/101080/0140239020201759038 . 2020b “Artificial Intelligence, Drone Swarming and Escalation Risks in Future Warfare.” The RUSI Journal 165 (2): 26–36

https://doi.org/101080/0307184720201752026 . 2020c “Artificial Intelligence in Nuclear Warfare: A Perfect Storm of Instability?” Washington Quarterly 43 (2): 197–211. https://doi.org/101080/0163660X20201770968 . 2022a “Automating the OODA Loop in the Age of Intelligent Machines: Reaffirming the Role of Humans in Command-and-Control DecisionMaking in the Digital Age.” Defence Studies 23 (1): 43–67 https://doi.org/101080/1470243620222102486 . 2022b “The AI Commander Problem: Ethical, Political, and Psychological Dilemmas of Human-Machine Interactions in AI-Enabled Warfare.” Journal of Military Ethics 21 (3–4): 246–71 EXPLORING DECISION ADVANT AGES | 287 https://doi.org/101080/1502757020232175887 Joint Chiefs of Staff. 2019 Joint Fire Support, Joint Publication 3-09 https://www.jcsmil/Portals/36/Documents/Doctrine/pubs/jp3 09pdf Jordan, David, James D. Kiras, David J Lonsdale, Christopher Tuck, and C Dale Walton. 2016 Understanding Modern Warfare: Second Edition

Understanding Modern Warfare: Second Edition. Second edi Cambridge: Cambridge University Press. https://doi.org/101017/CBO9781316460276 Kamrani, Farzad, Linus Kanestad, Linus J Luotsinen, and Björn Pelzer. 2023 “Attacking and Deceiving Military AI Systems.” Kania, Elsa. 2019 “Artificial Intelligence in Future Chinese Command Decision Making.” Artificial Intelligence , China , Russia , and the Global Order, 154– 61. Kania, Elsa B. 2020 “Minds at War: China’s Pursuit of Military Advantage through Cognitive Science and Biotechnology.” Prism 8 (3): 83–101 Kasampalis, Stylianos, Pavlos I Lazaridis, Zaharias D Zaharis, Aristotelis Bizopoulos, Spyridon Zettas, and John Cosmas. 2013 “Comparison of ITM and ITWOM Propagation Models for DVB-T Coverage Prediction.” In 2013 IEEE International Symposium on Broadband Multimedia Systems and Broadcasting (BMSB), 1–4. https://doi.org/101109/BMSB20136621780 Keaney, Thomas A., and Eliot A Cohen 1993 “Gulf War Air Power Survey

Summary Report.” Washington, D C Khan, Fazle Rabby, Md Muhabullah, Roksana Islam, Mohammad Monirujjaman Khan, Mehedi Masud, Sultan Aljahdali, Avinash Kaur, and Parminder Singh. 2021 “A Cost-Efficient Autonomous Air Defense System for National Security.” Security and Communication Networks 2021 https://doi.org/101155/2021/9984453 Kline, Alexander, Darryl Ahner, and Raymond Hill. 2019 “The Weapon-Target Assignment Problem.” Computers and Operations Research Elsevier Ltd https://doi.org/101016/jcor201810015 Kobzan, Sergiy. 2022 “GIS FOR THE ARMED FORCES OF UKRAINE TWO COMPONENTS Of VICTORY.” InterConf, July, 347–53 https://doi.org/1051582/interconf19-20072022037 Kott, Alexander, and David S. Alberts 2017 “How Do You Command an Army of Intelligent Things?” Computer 50 (12): 96–100. https://doi.org/101109/MC20174451205 Kuhn, H.W 1955 “The Hungarian Method for the Assignment Problem†” Naval Research Logistics Quarterly, no. March

https://doi.org/101002/nav3800020109 Kuhn, Thomas. 2020 The Structure of Scientific Revolutions Edited by Otto Neurath. International Encyclopedia of Unified Science Vol Foundation Chicago: The University of Chicago Press. https://doi.org/105840/philstudies196413082 Lai, Chyh Ming, and Tsung Hua Wu. 2019 “Simplified Swarm Optimization with Initialization Scheme for Dynamic Weapon–Target Assignment Problem.” Applied Soft Computing Journal 82 (September) 288 | EXPLORING DECISION ADVANT AGES https://doi.org/101016/jasoc2019105542 Lande, Dmytro, Igor Subach, and Alexander Puchkov. 2020 “A System for Analysis of Big Data from Social Media.” Information & Security: An International Journal 47 (1): 44–61. https://doiorg/1011610/ISIJ4703 Langley, Pat. 2012 “Intelligent Behavior in Humans and Machines” Advances in Cognitive Systems 2: 3–12. wwwaaaiorg Law, Averill M. 2002 “HOW TO BUILD VALID AND CREDIBLE SIMULATION MODELS.” In Proceedings of the 2022 Winter

Simulation Conference, edited by And B. Feng, G Pedrielli, Y Peng, S Shashaani, E Song, CG Corlu, L.H Lee, EP Chew, T Roeder and Eds P Lendermann, 1283–95 IEEE. https://doiorg/105555/35862103586317 Law, John. 1992 “Notes on the Theory of the Actor-Network: Ordering, Strategy, and Heterogeneity.” Systems Practice 5 (4): 379–93 https://doi.org/101007/BF01059830 Lawson, Stephanie. 2015 Theories of International Relations, Contending Approaches to World Politics. Theories of International Relations, Contending Approaches to World Politics. Cambridge: Polity Press https://www.absedugh/wp-content/uploads/2022/10/Sup-textbookTheories-of-International-Relationspdf LeCun, Yann. 2018 “The Power and Limits of Deep Learning” Article Research Technology Management 61 (6): 22–27. https://doi.org/101080/0895630820181516928 Lecun, Yann, Yoshua Bengio, and Geoffrey Hinton. 2015 “Deep Learning” Nature 521 (7553): 436–44. https://doiorg/101038/nature14539 Lemay Center for Doctrine.

2020 A Primer on Doctrine Lewin, Timothy P. 2019 “Targeting in Multi-Domain Operations: A Proposal to Update the U.S Army’s Targeting Process for the Modern Operating Environment.” Small Wars Journal https://smallwarsjournal.com/jrnl/art/targeting-multi-domainoperations-proposal-update-us-armys-targeting-process-modern Li, Guangjian, Guangjun He, Mingfa Zheng, and Aoyu Zheng. 2023 “Uncertain Sensor–Weapon–Target Allocation Problem Based on Uncertainty Theory.” Symmetry 15 (1) https://doiorg/103390/SYM15010176 Lindsay, Jon R, and Erik Gartzke. 2020 “Politics by Many Other Means: The Comparative Strategic Advantages of Operational Domains.” https://doi.org/101080/0140239020201768372 Liu, Hin Yan. 2019 “From the Autonomy Framework towards Networks and Systems Approaches for ‘Autonomous’ Weapons Systems.” Journal of International Humanitarian Legal Studies 10 (1). https://doi.org/101163/18781527-01001010 Loscher, Georg, Violetta Splitter, and David Seidl. 2019

“Theodore Schatzki’s Practice Theory and Its Implications for Organization Studies.” In Management, Organizations and Contemporary Social Theory, 115–34. Routledge. https://doiorg/104324/9780429279591-7 Lu, Yiping, and Danny Z. Chen 2021 “A New Exact Algorithm for the WeaponTarget Assignment Problem” Omega (United Kingdom) 98 (January): 1– 15. https://doiorg/101016/jomega2019102138 EXPLORING DECISION ADVANT AGES | 289 Lundgren, Jan, Mikael Rönnqvist, and Peter Värbrand. 2010 Optimization Lund: Studentlitteratur AB. Luotsinen, Linus J, Daniel Oskarsson, Peter Svenmarck, and Ulrika Wickenberg Bolin. 2019 “Explainable Artificial Intelligence: Exploring XAI Techniques in Military Deep Learning Applications.” Ma, Yingying, Guoqiang Wang, Xiaoxuan Hu, and He Luo. 2021 “Two-Stage Hybrid Heuristic Search Algorithm for Novel Weapon Target Assignment Problems.” Computers and Industrial Engineering 162 (December) https://doi.org/101016/jcie2021107717 Manne, Alan S.

1958 “A Target-Assignment Problem” Operations Research 6 (3): 346–51. https://doiorg/101287/OPRE63346 Matiuzzi Stocchero, Jorgito, Andre Dexheimer Carneiro, Iulisloi Zacarias, and Edison Pignaton de Freitas. 2023 “Combining Information Centric and Software Defined Networking to Support Command and Control Agility in Military Mobile Networks.” Peer-to-Peer Networking and Applications 16 (2): 765–84. https://doiorg/101007/s12083-022-01443-z Mattis, Jim, and Bing West. 2019 Call Sign Chaos New York, NY, USA: Random House. Mccann, Carol, and Ross Pigeau. 1999 “Clarifying the Concepts of Control and of Command.” Proceedings of the 5 Th Command and Control Research and Technology Symposium. Newport, RI: Navy War College, 16 Mcgrath, By James R. 2016 “Twenty-First Century Information Warfare and the Third Offset Strategy.” Joint Force Quarterly 82: 16–23 https://ndupress.nduedu/Portals/68/Documents/jfq/jfq-82/jfq-82 1623 McGrathpdf%0Ahttp://ndupressnduedu/JFQ/ McKernan,

Bethan, and Harry Davies. 2024 “‘The Machine Did It Coldly’: Israel Used AI to Identify 37,000 Hamas Targets.” The Guardian, April 3, 2024 https://www.theguardiancom/world/2024/apr/03/israel-gaza-aidatabase-hamas-airstrikes Meerveld, H. W, R HA Lindelauf, E O Postma, and M Postma 2023 “The Irresponsibility of Not Using AI in the Military.” Ethics and Information Technology 25 (1): 1–6. https://doiorg/101007/s10676-023-09683-0 Metler, William A, Fred L Preston, Jim Hofmann, F L Preston, and James B Hofmann. 1990 “Nq A Suite of Weapon Assignment Algorithms N for a SDI Mid-Course Battle Manager A Suite of Weapon Assignment Algorithms for a SDI Mid-Course Battle Manager.” Minsky, Marvin. 1986 The Society of Mind New York, USA: Simon & Schuster Mintzberg, Henry, B Ahlstrand, and Joseph Lampel. 2009 Strategy Safari: A Guided Tour Through The Wilds of Strategic Management. 2nd editio Vol 1998. FT Press Mojtahed, Vahid, Ulrika Wickenberg Bolin, Ola Svenonius, and Magnus

Pierrau. 2023. “Mot En Datadriven Försvarsmakt Mot En Datadriven Försvarsmakt Slutrapport För Studien Informationshantering i Försvarsmakten.” Mol, Annemarie. 2010 “Actor-Network Theory: Sensitive Terms and Enduring Tensions.” Kölner Zeitschrift Für Soziologie Und Sozialpsychologie Sonderheft Vol. 50 (No 1): 253–69 https://dareuvanl 290 | EXPLORING DECISION ADVANT AGES Monte, Louis A Del. 2018 “Genius Weapons : Artificial Intelligence, Autonomous Weaponry, and the Future of Warfare.” Prometheus Books 2019: 1–2. Nash, John C. 2000 “The (Dantzig) Simplex Method for Linear Programming” Computing in Science and Engineering 2 (1): 29–31. https://doi.org/101109/5992814654 National Security Commission. 2021 “Final Report National Security Commission on Artificial Intelligence.” NATO. 2016 “AJP-39 Allied Joint Doctrine for Joint Targeting” NATO Standardization Office, no. April: 2-4,2-5 https://assets.publishingservicegovuk/government/uploads/system/u

ploads/attachment data/file/628215/20160505nato targeting ajp 3 9.pdf%0Ahttps://wwwgovuk/government/upload s/system/uploads/attachment data/file/628215/20160505nato targeting ajp 3 9.p . 2019 “Allied Joint Doctrine for the Conduct of Operations, AJP-3,” no February: 164. https://wwwgovuk/government/publications/alliedjoint-doctrine-for-the-conduct-of-operations-ajp-3b . 2020 “Science and Technology Trends 2020-2040” Nato Science and Technology Organisation, 160. http:%5Cwwwstonatoint NATO Science & Technology Organization. 2023 “NATO Science & Technology Trends 2023-2043 Stt23-Vol 1.” Nguyen, Thanh Thi, Ngoc Duy Nguyen, and Saeid Nahavandi. 2020 “Deep Reinforcement Learning for Multiagent Systems: A Review of Challenges, Solutions, and Applications.” IEEE Transactions on Cybernetics 50 (9): 3826–39. https://doiorg/101109/TCYB20202977374 NSO. 2021 Allied Joint Doctrine for Joint Targeting Nato Standard Ajp-39 Vol Edition B. Nurkin, Tate, and Julia Siegel. 2023

“Battlefield Applications for HumanMachine Teaming : Demonstrating Value, Experimenting with New Capabilities and Accelerating Adoption.” https://www.atlanticcouncilorg/wpcontent/uploads/2023/08/Battlefield-Applications-for-HMTpdf O’Dowd, Edward, and Arthur Waldron. 1991 “Sun Tzu for Strategists” Comparative Strategy 10 (1): 25–36. https://doi.org/101080/01495939108402828 Office of the Chairman of the Joint Chiefs of Staff. 2020 DOD Dictionary of Military and Associated Terms. US, Department of Defense https://www.jcsmil/Portals/36/Documents/Doctrine/pubs/dictionary pdf. Olsen, John Andreas. 2017 “Warden Revisited - The Pursuit of Victory through Air Power.” Air Power History 64 (4 (winter 2017)): 39–53 . , ed 2023 Airpower Pioneers: From Billy Mitchell to Dave Deptula Annapolis, Maryland: Naval Institute Press. Oniani, David, Jordan Hilsman, Yifan Peng, Ronald K. Poropatich, Jeremy C Pamplin, Gary L. Legault, and Yanshan Wang 2023 “Adopting and Expanding Ethical

Principles for Generative Artificial Intelligence from EXPLORING DECISION ADVANT AGES | 291 Military to Healthcare.” Npj Digital Medicine 6 (1): 1–23 https://doi.org/101038/s41746-023-00965-x Osinga, Frans P.B 2007 Science, Strategy and War: The Strategic Theory of John Boyd. Science, Strategy and War: The Strategic Theory of John Boyd Abingdon: Routledge. https://doiorg/104324/9780203088869 Osinga, Frans P.B, and Mark P Roorda 2015 “From Douhet to Drones, Air Warfare, and the Evolution of Targeting.” In Targeting: The Challenges of Modern Warfare, edited by Paul A.L Ducheine, Michael N Schmitt, and Frans P.B Osinga, 27–76 The Hague: TMC Asser Press https://doi.org/101007/978-94-6265-072-5 Pape, Robert A. 2014 Bombing to Win: Air Power and Coercion in War Cornell University Press. https://doiorg/doi:107591/9780801471513 Park, Hyeon Woo, and Han Lim Choi. 2023 “Weapon-Target Assignment and Firing Scheduling for Rapid Engagement with Heterogeneous Interceptors.”

International Journal of Aeronautical and Space Sciences, July. https://doiorg/101007/S42405-023-00572-W Payne, Kenneth. 2018 Strategy, Evolution, and War From Apes to Artificial Intelligence. Georgetown University Press Penney, Heather R. 2023 “Scale, Scope, Speed & Survivability:Winning the Kill Chain Competition.” Mitchel Institute 40 (Policy Paper): 1–40 Pickering, Andrew. 2005 “Practice and Posthumanism: Social Theory and a History of Agency.” The Practice Turn in Contemporary Theory, January, 172–83. https://doiorg/104324/9780203977453-21/PRACTICEPOSTHUMANISM Pinnix, Curtis E. 2021 “Specialized Analytic and Targeting Study: A Methodology and Approach for Conducting Faster Full-Spectrum Targeting.” Joint Force Quarterly 103 (4th Quarter): 72–78 Polit, Denise F., and Cheryl Tatano Beck 2010 “Generalization in Quantitative and Qualitative Research: Myths and Strategies.” International Journal of Nursing Studies 47 (11): 1451–58.

https://doi.org/101016/jijnurstu201006004 Pratzner, Phillip R. 2015 “The Current Targeting Process” In Targeting: The Challenges of Modern Warfare, edited by Paul A.L Ducheine, Michael N Schmitt, and Frans P.B Osinga, 77–97 The Hague: TMC Asser Press Qiao-Franco, Guangyu, and Ingvild Bode. 2023 “Weaponised Artificial Intelligence and Chinese Practices of Human–Machine Interaction.” Chinese Journal of International Politics 16 (1): 106–28. https://doi.org/101093/cjip/poac024 Quttineh, Nils Hassan, and Torbjörn Larsson. 2014 “Military Aircraft Mission Planning: Efficient Model-Based Metaheuristic Approaches.” Optimization Letters 9 (8): 1625–39. https://doiorg/101007/s11590-014-0831-x Quttineh, Nils Hassan, Torbjörn Larsson, Kristian Lundberg, and Kaj Holmberg. 2013. “Military Aircraft Mission Planning: A Generalized Vehicle Routing Model with Synchronization and Precedence.” EURO Journal on Transportation and Logistics 2 (1–2): 109–27.

https://doi.org/101007/s13676-013-0023-3 Radder, Hans. 2009 “The Philosophy of Scientific Experimentation: A Review” 292 | EXPLORING DECISION ADVANT AGES Automated Experimentation 1 (1): 2. https://doiorg/101186/17594499-1-2 Rahman, Md Atiqur, and Yang Wang. 2016 “Optimizing Intersection-OverUnion in Deep Neural Networks for Image Segmentation” In Advances in Visual Computing, edited by George Bebis, Richard Boyle, Bahram Parvin, Darko Koracin, Fatih Porikli, Sandra Skaff, Alireza Entezari, et al., 234–44 Cham: Springer International Publishing. Rajabi, Zeinab. 2012 “Data-Centric Enterprise Architecture,” no August: 53– 60. https://doiorg/105815/ijieeb20120408 Richards, Chet. 2011 “Boyd’s OODA Loop” Proceedings of Lean Software and Systems Conference 2011, 127–36. https://sslserver.com/jvminccom/boydsrealooda looppdf . 2012 “Boyd ’ s OODA Loop” 5 (1): 127–36 . 2020 “Boyd ’ s OODA Loop” Necesse, Sjokrigsskolen 5 (1): 142–64

https://fhs.brageunitno/fhsxmlui/bitstream/handle/11250/2683228/Boyds OODA Loop Necesse vol 5 nr 1.pdf?sequence=1&isAllowed=y Ritchey, Tom. 1991 “Analysis and Synthesis: On Scientific Method – Based on a Study by Bernhard Riemann.” Systems Research 8 (4): 21–41 https://doi.org/101002/sres3850080402 Ritchie, Jane, Jane Lewis, Carol McNaughton Nicholls, and Rachel Ormston, eds. 2014. Qualitative Research Practice: A Guide for Social Science Students and Researchers. Second edition London: SAGE Publications Ltd Robinson, Stephen. 2021 The Blind Strategist Dunedin, NZ: Exilse Publishing Pty Ltd. Robinson, Stewart. 2017 “A TUTORIAL ON SIMULATION CONCEPTUAL MODELING.” In Proceedings - Winter Simulation Conference, edited by Eds. W K V Chan, A D’Ambrogio, G Zacharewicz, N Mustafee, G Wainer, and E. Page, 2740–51 https://doiorg/105555/32421813242222 Roff, Heather. 2020 “AI Deception: When Your Artificial Intelligence Learns to Lie.” IEEE Spectrum 2020

https://spectrum.ieeeorg/automaton/artificial-intelligence/embeddedai/ai-deception-when-your-ai-learns-to-lie Ronneberger, Olaf, Philipp Fischer, and Thomas Brox. 2015 “U-Net: Convolutional Networks for Biomedical Image Segmentation.” In Medical Image Computing and Computer-Assisted Intervention -- MICCAI 2015, edited by Nassir Navab, Joachim Hornegger, William M Wells, and Alejandro F Frangi, 234–41. Cham: Springer International Publishing Rumsfeld, Donald H. 2002 “Transforming the Military” Foreign Affairs 81 (3): 20–32. https://doiorg/102307/20033160 Russell, Stuart. 2019 Human Compatible: Artificial Intelligence and the Problem of Control. Book [New York, NY] : Viking, https://search.proquestcom/docview/2424915021?accountid=41712 . 2023 “AI Weapons: Russia’s War in Ukraine Shows Why the World Must Enact a Ban.” Nature 2023 614:7949 614 (7949): 620–23 https://doi.org/101038/d41586-023-00511-5 Russell, Stuart, Sabine Hauert, Russ Altman, and Manuela Veloso. 2015

EXPLORING DECISION ADVANT AGES | 293 “Robotics: Ethics of Artificial Intelligence.” Nature 521 (7553): 415–18 https://doi.org/101038/521415a Russell, Stuart, and Peter Norvig. 2014 Artificial Intelligence A Modern Approach Third Edition. Book Pearson Pearson Education Limited https://doi.org/101017/S0269888900007724 Rynning, Sten, Olivier Schmitt, and Amelie Theussen, eds. 2021 War Time Washington, D.C: Brookings Institution Press Saralioglu, Ekrem, and Oguz Gungor. 2022 “Semantic Segmentation of Land Cover from High Resolution Multispectral Satellite Images by SpectralSpatial Convolutional Neural Network.” Geocarto International 37 (2): 657–77. https://doiorg/101080/1010604920201734871 Sarker, Iqbal H. 2021 “Deep Learning: A Comprehensive Overview on Techniques, Taxonomy, Applications and Research Directions.” SN Computer Science 2 (6): 1–20. https://doiorg/101007/s42979-02100815-1 Sarker, Iqbal H., Helge Janicke, Mohamed Amine Ferrag, and Alsharif Abuadbba. 2024

“Multi-Aspect Rule-Based AI: Methods, Taxonomy, Challenges and Directions towards Automation, Intelligence and Transparent Cybersecurity Modeling for Critical Infrastructures.” Internet of Things (Netherlands) 25 (September 2023): 101110. https://doi.org/101016/jiot2024101110 Sawyer, Ralph D. 1994 Sun Tzu: The Art of War Basic Books Scharre, Paul. 2018a Army of None New York, USA: W W Norton & Company . 2018b “How Swarming Will Change Warfare” Bulletin of the Atomic Scientists 74 (6). https://doiorg/101080/0096340220181533209 . 2019 Framtidens Fjärrstyrda Krig: Soldatrobotar, Drönare och Artificiell Intelligens. Book Edited by Mikael Dahlgren Stockholm: Svenskt militärhistoriskt biblioteks förlag. Schatzki, Theodore. 2013 “The Edge of Change: On the Emergence, Persistence and Dissolution of Practices.” Sustainable Practices, March, 47–62 https://doi.org/104324/9780203071052-9 Schatzki, Theodore R. 2012 A Primer on Practices: Theory and Research Edited by Joy

Higgs, Ronald Barnett, Stephen Billett, Maggie Hutchings, and Franziska Trede. Brill https://brill.com/view/book/edcoll/9789462091283/BP000003xml Schmitt, Olivier. 2021 “Speed, Time, and Western Military Power” In Wartime - Temporality and the Decline of Western Military Power, edited by Sten Rynning, Olivier Schmitt, and Amelie Theussen, 316. Washington, D C: Brookings Institution Press. Schubert, Johan, Joel Brynielsson, Mattias Nilsson, and Peter Svenmarck. 2018 “Artificial Intelligence for Decision Support in Command and Control Systems Topic Topic 9: Knowledge Systems Artificial Intelligence for Decision Support in Command and Control Systems.” . 2019 “Artificial Intelligence for Decision Support in Command and Control Systems.” 23rd International Command and Control Research & Technology Symposium, no. November 2018: 15–30 Segrè, Claudio G. 1992 “Giulio Douhet: Strategist, Theorist, Prophet?” Journal 294 | EXPLORING DECISION ADVANT AGES of Strategic

Studies 15 (3): 351–66. https://doi.org/101080/01402399208437489 Shadish, W. R, Cook, T D, Campbell, DT 2002 Experimental and Designs for Generalized Causal Inference. Experimental and Quasi-Experimental Design for Causual Inference. Shapere, Dudley. 1984 “The Character of Scientific Change” In Reason and the Search for Knowledge: Investigations in the Philosophy of Science, 205–60. Dordrecht: Springer Netherlands. https://doiorg/101007/978-94-0109731-4 11 Shapiro, Susan P. 2005 “Agency Theory” Annual Review of Sociology https://doi.org/101146/annurevsoc31041304122159 Silav, Ahmet, Esra Karasakal, and Orhan Karasakal. 2022 “Bi-Objective Dynamic Weapon-Target Assignment Problem with Stability Measure.” Annals of Operations Research 311 (2): 1229–47. https://doi.org/101007/s10479-020-03919-8 Skolnik, Merrill I. 2001 Introduction to Radar Systems Third edit New York: McGraw-Hill Education,. Soare, Simona R., Pavneet Singh, and Meia Nouwens 2023 “Software-Defined

Defence: Algorithms at War.” https://wwwiissorg/globalassets/medialibrary---content--migration/files/research-papers/iiss softwaredefined-defence 17022023pdf Stavridis, James, and Elliot Ackerman. 2024 “Drone Swarms Are About to Change the Balance of Military Power.” The Wall Street Journal, 2024 https://www.wsjcom/tech/drone-swarms-are-about-to-change-thebalance-of-military-power-e091aa6f Steen, Marc, Jurriaan van Diggelen, Tjerk Timan, and Nanda van der Stap. 2023 “Meaningful Human Control of Drones: Exploring Human–Machine Teaming, Informed by Four Different Ethical Perspectives.” AI and Ethics 3 (1): 281–93. https://doiorg/101007/s43681-022-00168-2 Stiegler, Bernard. 1998 Technics and Time 1 The Fault of Epimetheus Stanford: Stanford University Press. Svenmarck, Peter, Linus Luotsinen, Mattias Nilsson, and Johan Schubert. 2018a. “Possibilities and Challenges for Artificial Intelligence in Military Applications.” Proceedings of the NATO Big Data and Artificial

Intelligence for Military Decision Making Specialists’ Meeting, 1–17. https://www.researchgatenet/publication/326774966 . 2018b “Possibilities and Challenges for Artificial Intelligence in Military Applications.” Proceedings of the NATO Big Data and Artificial Intelligence for Military Decision Making Specialists’ Meeting, 1–17. https://www.researchgatenet/publication/326774966 Swedberg, Richard. 2012 “Theorizing in Sociology and Social Science: Turning to the Context of Discovery.” Theory and Society 41 (1): 1–40 https://doi.org/101007/s11186-011-9161-5 . , ed 2014 Theorizing in Social Science: The Context of Discovery California, USA: Stanford University Press. Szeligowski, Robert. 2018 “NAVAL WAR COLLEGE Newport, RI COGNIFYING THE OODA LOOP: IMPROVED MARITIME DECISION MAKING” 298 EXPLORING DECISION ADVANT AGES | 295 (0704). Tarraf, Danielle, William Shelton, Edward Parker, Brien Alkire, Diana Carew, Justin Grana, Alexis Levedahl, et al. 2019 The Department

of Defense Posture for Artificial Intelligence: Assessment and Recommendations. The Department of Defense Posture for Artificial Intelligence: Assessment and Recommendations. RAND Corporation https://doiorg/107249/rr4229 Taye, Muhammad Mustafa. 2023 “Understanding of Machine Learning with Deep Learning :” Computers MDPI 12 (91): 1–26. Teigen, Jeremy M. 2014 “Experimental Methods in Military and Veteran Studies.” Routledge Handbook of Research Methods in Military Studies, 228–37. https://doiorg/104324/9780203093801ch20 Tversky, Amos, and Daniel Kahneman. 1974 “Judgment under Uncertainty: Heuristics and Biases” 185 (4157): 1124–31. U.S Joint Chiefs of Staff 2019 Joint Air Operations U.S Joint Targeting School 2017 “Joint Targeting School Student Guide” Dam Neck, Virginia. UK Ministry of Defense. 2016 J9 FOI Secretariat Permanent Joint Headquarters Northwood Headquarters Sandy Lane United Kingdom. Vol 1 . 2023 “Joint Doctrine Note 1/23, Intelligence, Surveillance

and Reconnaissance.” UK Ministry of Defense https://doi.org/104324/9781315650463-23 UK MoD. 2018 “Global Strategic Trends The Future Starts Today Sixth Edition.” . 2021 “Defence Experimentation for Force Development Handbook” https://assets.publishingservicegovuk/government/uploads/system/u ploads/attachment data/file/957195/20210121DEFD Handbook Version 2-O.pdf US. 2020 CJCSI 512002E Joint Doctrine Development System https://www.jcsmil/Library/CJCSInstructions/u46626q/353132302E3032/u46626c/616C6C636F6C756 D6E73/u46626m/contain/. US Air Force. 2021 Air Force Doctrine Publication 3-60, Targeting www.doctrineafmil/Doctrine-Publications/AFDP-3-60-Targeting US Chairman Joint Chief of Staff. 2021 CJCSM 310801 Joint Fires Element www.jcsmil/library . 2022 CJCSI 350501E US Congress. 1986 “Goldwater-Nichols Department of Defense Reorganization Act of 1986.” https://doiorg/101016/0022-460X(79)90914-3 US Department of Defense. 2018 FY18 AIR FORCE PROGRAMS Small Diameter Bomb ( SDB

) II. . 2019 Experimentation Guidebook https://dod.defensegov/Portals/1/Documents/pubs/2018-NationalDefense-Strategy-Summarypdf . 2023 DOD Directive 300009 AUTONOMY IN WEAPON SYSTEMS US DoD. 2016a “CJCSI 337001B 6 May 2016 TARGET DEVELOPMENT STANDARDS.” . 2016b “Target Development Standards CJCSI 337001,” no May 296 | EXPLORING DECISION ADVANTAGES . 2018 “Summary of the 2018 Department of Defense Artificial Intelligence Strategy,” 1–17. https://media.defensegov/2019/Feb/12/2002088963/-1/1/1/SUMMARY-OF-DOD-AI-STRATEGYPDF . 2020 “Ethical Principles for Artificial Intelligence and Data Analytics” US DoD ADP 1-01. 2019 “Doctrine Primer (ADP 1-01),” no JULY: Glossary-1 https://books.googlecom/books?id=WiXHDwAAQBAJ US DoD Defense Science Board. 2016a “DSB Summer Study on Autonomy” . 2016b “Final Report of the Defense Science Board Summer Study on Autonomy.” https://www.researchgatenet/publication/306286423 Final Report of the Defense Science Board Summer

Study on Autonomy PubliclyReleasable Version. US DoD Joint Publication. 2018a Joint Publication 3-0, Joint Operations Joint Publication 3-0. https://www.jcsmil/Portals/36/Documents/Doctrine/pubs/jp3 0ch1p df. . 2018b Joint Publication 3-60, Joint Targeting USA Vassiliou, Marius S., David S Alberts, and Jonathan Russell Agre 2015 C2 ReEnvisioned: The Future of the Enterprise - 1st Edition CRC Press https://www.routledgecom/C2-Re-envisioned-The-Future-of-theEnterprise/Vassiliou-Alberts-Agre/p/book/9781466595804 Verdiesen, Ilse, Filippo Santoni De Sio, and Virginia Dignum. 2021 “Accountability and Control Over Autonomous Weapon Systems: A Framework for Comprehensive Human Oversight” 31: 137–63. https://doi.org/101007/s11023-020-09532-9 Vinci, Anthony. 2019 “Win-Win: A Public Private Strategy for Dramatically Improving National Security Tech.” Virilio, Paul. 2002 “Desert Screen: War at the Speed of Light” i (2001): 148

https://www.amazoncom/Desert-Screen-Speed-ContinuumImpacts/dp/0826479340/ref=sr 1 37?crid=3L442AU7XU0OR&dchild=1 &keywords=paul+virilio&qid=1603392604&s=books&sprefix=pail%2Cs tripbooks-intl-ship%2C248&sr=1-37. Wang, Haiying, and Fang Miao. 2022 “Building Extraction from Remote Sensing Images Using Deep Residual U-Net.” European Journal of Remote Sensing 55 (1): 71–85. https://doi.org/101080/2279725420212018944 Wang, Tong, Liyue Fu, Zhengxian Wei, Yuhu Zhou, and Shan Gao. 2023 “Unmanned Ground Weapon Target Assignment Based on Deep QLearning Network with an Improved Multi-Objective Artificial Bee Colony Algorithm.” Engineering Applications of Artificial Intelligence 117 (January). https://doiorg/101016/jengappai2022105612 Wang, Weiyun, Zhe Chen, Wenhai Wang, Yue Cao, Yangzhou Liu, Zhangwei Gao, Jinguo Zhu, et al. 2024 “Enhancing the Reasoning Ability of Multimodal Large Language Models via Mixed Preference Optimization.”

http://arxiv.org/abs/241110442 Wang, X. W, and F Wang 2020 “THE PRECISION of GOOGLE EARTH MAP EXPLORING DECISION ADVANT AGES | 297 ANALYSIS with the COORDINATES of IGS STATIONS.” International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences - ISPRS Archives 42 (3/W10): 1053–56. https://doi.org/105194/isprs-archives-XLII-3-W10-1053-2020 Wang, Yipeng, Bin Xin, and Jie Chen. 2022 “An Adaptive Memetic Algorithm for the Joint Allocation of Heterogeneous Stochastic Resources.” IEEE Transactions on Cybernetics 52 (11): 11526–38. https://doi.org/101109/TCYB20213087363 Watson, Michael, and Peter Cacioppi. 2014 “CPLEX Network-DesignProgramming” http://petercacioppicom/ Webster, Murray, and Jane Sell, eds. 2014 Laboratory Experiments in the Social Sciences: Second Edition. Laboratory Experiments in the Social Sciences: Second Edition. 2nd editio Academic Press Inc https://doi.org/101016/C2011-0-07562-2 Welsh, Jennifer M. 2015 “The Morality of

‘Drone Warfare’” In Drones and the Future of Armed Conflict : Ethical, Legal, and Strategic Implications. Williams, Paul H. 2013 Model Building in Mathematical Programming 5th editio. Chichester: John Wiley & Sons, Ltd Wong, Yuna, John Yurchak, Robert Button, Aaron Frank, Burgess Laird, Osonde Osoba, Randall Steeb, Benjamin Harris, and Sebastian Bae. 2020 Deterrence in the Age of Thinking Machines. Deterrence in the Age of Thinking Machines. RAND Corporation https://doiorg/107249/rr2797 Xu, Yongjun, Xin Liu, Xin Cao, Changping Huang, Enke Liu, Sen Qian, Xingchen Liu, et al. 2021 “Artificial Intelligence: A Powerful Paradigm for Scientific Research.” Innovation 2 (4) https://doi.org/101016/jxinn2021100179 Yan, Guilong. 2020 “The Impact of Artificial Intelligence on Hybrid Warfare” Small Wars and Insurgencies 31 (4): 898–917. https://doi.org/101080/0959231820191682908 Yang, Li, Yifei Zou, Dongxiao Yu, and Jiguo Yu. 2023 “Distributed Age-ofInformation Optimization in

Edge Computing for Internet of Vehicles” Journal of Systems Architecture 144 (November): 103000. https://doi.org/101016/JSYSARC2023103000 Yates, Roy D., Yin Sun, D Richard Brown, Sanjit K Kaul, Eytan Modiano, and Sennur Ulukus. 2021 “Age of Information: An Introduction and Survey” IEEE Journal on Selected Areas in Communications 39 (5): 1183–1210. https://doi.org/101109/JSAC20213065072 Zellen, Barry. 2015 “Bernard Brodie and the Bomb: At the Birth of the Bipolar World.” University of Lapland Zhao, Zhongliang, Pedro Cumino, Arnaldo Souza, Denis Rosário, Torsten Braun, Eduardo Cerqueira, and Mario Gerla. 2019 “Software-Defined Unmanned Aerial Vehicles Networking for Video Dissemination Services.” Ad Hoc Networks 83 (February): 68–77. https://doi.org/101016/JADHOC201808023 Zhou, Longyu, Supeng Leng, Qiang Liu, Haoye Chai, and Jihua Zhou. 2021 “Intelligent Sensing Scheduling for Mobile Target Tracking Wireless Sensor Networks.” GENERIC COLORIZED JOURNAL, 1 298 |

EXPLORING DECISION ADVANT AGES Zinca, Diana-Ioana. 2018 “JOINT TARGETING SYSTEM – THE SUPPORT TOOL FOR JOINT TARGETING CYCLE AND ITS APLICABILITY IN COMPUTER ASSISSTED EXERCISES.” National Defense University EXPLORING DECISION ADVANT AGES | 299