Close
About
FAQ
Home
Collections
Login
USC Login
Register
0
Selected
Invert selection
Deselect all
Deselect all
Click here to refresh results
Click here to refresh results
USC
/
Digital Library
/
University of Southern California Dissertations and Theses
/
The human element: addressing human adversaries in security domains
(USC Thesis Other)
The human element: addressing human adversaries in security domains
PDF
Download
Share
Open document
Flip pages
Contact Us
Contact Us
Copy asset link
Request this asset
Transcript (if available)
Content
The Human Element: Addressing Human Adversaries in Security Domains by James Pita A Dissertation Presented to the FACULTY OF THE USC GRADUATE SCHOOL UNIVERSITY OF SOUTHERN CALIFORNIA In Partial Fulfillment of the Requirements for the Degree DOCTOR OF PHILOSOPHY (COMPUTER SCIENCE) December 2012 Copyright 2012 James Pita Acknowledgments An individual and special thank you belongs to my adviser Milind Tambe. I cannot begin to thank you enough for your steadfast eort, determination, and dedication to each of your students. Not only are you the exemplar of what an adviser should be, but you are genuinely cherished by anyone who has had the pleasure to work with you. You have made this experience altogether remarkable due to your outstanding guidance and more importantly your sincere friendship. I would also like to thank my committee members for helping to guide my research and think beyond it: Jonathan Gratch, Richard John, Sarit Kraus, Stacy Marsella, and Nicholas Weller. I would particularly like to thank Sarit Kraus for her unparalleled insights, guidance, and assistance throughout my career and Richard John for helping me to expand my understanding of experi- mental approaches. Furthermore, I would like to thank my co-authors over the years: Bo An, Har- ish Bellamane, Shane Cullen, Manish Jain, Richard John, Christopher Kiekintveld, Sarit Kraus, Jun-young Kwak, Reuma Magori-Cohen, Rajiv Maheswaran, Janusz Marecki, Thanh Nguyen, Fernando Ord´ o˜ nez, Praveen Paruchuri, Christopher Portway, Shyamsunder Rathi, Michael Scott, Eric Shieh, Erin Steigerwald, Milind Tambe, Jason Tsai, Craig Western, Rong Yang, and Zhengyu Yin. Your dedicated eorts, assistance, guidance, and hard work made this experience exception- ally better. ii Beyond my mentors and collaborators, I would like to thank CREATE, the Los Angeles World Airport (LAWA) police, and the Transportation Security Administration for giving me the opportunity to work on real-world problems that have a direct impact on the community. A special thank you to Erroll Southers for tirelessly promoting ARMOR as a viable approach for critical security problems, Ernest Cruz for his countless hours developing the original ARMOR software package, and Shane Cullen and Erin Steigerwald for their eorts in developing the GUARDS system. I also want to thank my colleagues at USC and the greater TEAMCORE community. You have all helped make these past years a special experience for me and I will never forget all the times we have shared and all the help and guidance you have given me. A special thanks goes to Janusz Marecki for all the laughs and advice over the years. I would also like to thank God for this tremendous opportunity. Finally, more than thank you goes to my mother Diane Pita, father Eugene Pita, and brother Michael Pita. Thank you for a lifetime of support in everything and anything that I do. Without your unconditional love and support I would not have been able to get where I am today. Thank you for making me push my limits, explore the world around me, and expand my horizons. Thank you for all the sacrifices you have made to get me here and for your never ending guidance, eort, support, and love. Thank you for being the cornerstone of my life. iii Table of Contents Acknowledgments ii List of Figures vii Abstract ix Chapter 1: Introduction 1 1.1 Problem Addressed . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1 1.2 Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4 1.2.1 COBRA/MATCH . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5 1.2.2 Security Circumvention Games . . . . . . . . . . . . . . . . . . . . . . 8 1.3 Guide to Thesis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10 Chapter 2: Background 11 2.1 Stakelberg Games . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11 2.2 Bayesian Stackelberg Games . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13 2.3 Strong Stackelberg Equilibrium . . . . . . . . . . . . . . . . . . . . . . . . . . . 15 2.4 DOBSS and Baseline Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . 16 2.4.1 DOBSS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17 2.4.2 UNIFORM . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20 2.4.3 MAXIMIN . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21 2.5 BRQR . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21 2.6 Security Stackelberg Games . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23 2.7 Los Angeles International Airport . . . . . . . . . . . . . . . . . . . . . . . . . 26 2.8 Human Subjects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28 Chapter 3: Related Work 32 3.1 Computing Optimal Stackelberg Equilibria . . . . . . . . . . . . . . . . . . . . 32 3.1.1 Ecient Solutions to general Bayesian Stackelberg games . . . . . . . . 32 3.1.2 Ecient Solutions for Large-Scale Security Games . . . . . . . . . . . . 35 3.2 Computing Robust Strategies . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37 3.3 Addressing Suboptimal Decisions . . . . . . . . . . . . . . . . . . . . . . . . . 38 iv Chapter 4: COBRA Algorithm 42 4.1 Key Ideas . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44 4.1.1 Bounded Rationality . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44 4.1.2 Anchoring Theory . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45 4.2 Robust Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48 4.2.1 COBRA(0,) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48 4.2.2 COBRA(; 0) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50 4.2.3 COBRA(;) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51 4.2.4 Complexity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52 4.3 Equivalences Between Models . . . . . . . . . . . . . . . . . . . . . . . . . . . 53 4.4 Experiment Purpose, Design, and Results . . . . . . . . . . . . . . . . . . . . . 56 4.4.1 Purpose of this Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56 4.4.2 Experimental Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . 57 4.4.2.1 Participants . . . . . . . . . . . . . . . . . . . . . . . . . . . 58 4.4.2.2 Reward Structure . . . . . . . . . . . . . . . . . . . . . . . . 58 4.4.2.3 Observability Conditions . . . . . . . . . . . . . . . . . . . . 60 4.4.2.4 Algorithms and Parameters . . . . . . . . . . . . . . . . . . . 62 4.4.2.5 Experimental Procedure . . . . . . . . . . . . . . . . . . . . . 66 4.4.3 Experimental Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . 68 4.4.3.1 Key Observations . . . . . . . . . . . . . . . . . . . . . . . . 70 4.4.3.2 Statistical Significance . . . . . . . . . . . . . . . . . . . . . . 71 4.4.3.3 Analysis of Results . . . . . . . . . . . . . . . . . . . . . . . 74 4.4.4 Handling Observational Uncertainty . . . . . . . . . . . . . . . . . . . . 78 4.4.5 Runtime Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 85 Chapter 5: MATCH Algorithm 87 5.1 MATCH Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89 5.2 Experiment Purpose, Design, and Results . . . . . . . . . . . . . . . . . . . . . 95 5.2.1 Purpose of this Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95 5.2.2 Experimental Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96 5.2.2.1 Participants . . . . . . . . . . . . . . . . . . . . . . . . . . . 97 5.2.2.2 Reward Structure . . . . . . . . . . . . . . . . . . . . . . . . 98 5.2.2.3 Experimental Procedure . . . . . . . . . . . . . . . . . . . . . 100 5.2.3 Results for Original Structures . . . . . . . . . . . . . . . . . . . . . . . 103 5.2.4 Results for New Reward Structures . . . . . . . . . . . . . . . . . . . . 104 5.3 Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105 5.4 -Re-estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 108 5.5 Runtime Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 110 Chapter 6: Security Circumvention Games 112 6.1 TSA Security Challenges . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114 6.1.1 Modeling the TSA Resource Allocation Challenges . . . . . . . . . . . . 114 6.1.1.1 Defender Strategies . . . . . . . . . . . . . . . . . . . . . . . 115 6.1.1.2 Attacker Actions . . . . . . . . . . . . . . . . . . . . . . . . . 116 6.1.2 Compact Representation for Eciency . . . . . . . . . . . . . . . . . . 117 v 6.1.2.1 Threat Modeling for TSA . . . . . . . . . . . . . . . . . . . . 117 6.1.2.2 Compact Representation . . . . . . . . . . . . . . . . . . . . . 119 6.2 Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122 6.2.1 Security Policy Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . 122 6.2.2 Runtime Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 124 Chapter 7: Conclusions 127 7.1 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127 7.2 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130 Bibliography 133 Appendix A: Statistical Significance Tests 139 A.1 Statistical Significance Tests for COBRA . . . . . . . . . . . . . . . . . . . . . 139 A.2 Statistical Significance Tests for MATCH . . . . . . . . . . . . . . . . . . . . . 140 Appendix B: Reward Structures 141 Appendix C: Strategies 160 Appendix D: Expected Rewards for COBRA Experiments 197 Appendix E: Expected Response Percentages for COBRA Experiment 199 Appendix F: Strategies for varying in COBRA(,2.5) 200 Appendix G: Experimental Instructions 204 G.1 Material for COBRA Experiments . . . . . . . . . . . . . . . . . . . . . . . . . 204 G.2 Material for MATCH Experiments . . . . . . . . . . . . . . . . . . . . . . . . . 207 G.2.1 Obvious Games . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 207 G.3 Experiment Instructions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 208 vi List of Figures 2.1 LAX Security . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27 4.1 Game Interface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62 4.2 Single Observation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62 4.3 Average leader expected value . . . . . . . . . . . . . . . . . . . . . . . . . . . 69 4.4 Unobserved condition - Expected average reward . . . . . . . . . . . . . . . . . 80 4.5 Strategy entropy for varying values . . . . . . . . . . . . . . . . . . . . . . . . 81 4.6 Average expected values for varying under the unlimited observation condition 82 4.7 Comparing runtimes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 86 5.1 Game Interface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 96 5.2 1-Norm Scatter Plots . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99 5.3 Original reward structures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 103 5.4 Scatter Plot of Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 107 5.5 Re-estimated Reward Structures . . . . . . . . . . . . . . . . . . . . . . . . . . 110 5.6 Runtime results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 111 6.1 Policy Analysis: Increasing resources for 10 areas with 3 security activities per area123 6.2 X-axis: Areas, Y-axis: Runtime . . . . . . . . . . . . . . . . . . . . . . . . . . . 126 6.3 Runtime: Increasing resources for 10 areas with 3 security activities per area . . . 126 vii G.1 Game Interface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 204 G.2 Single Observation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 205 viii Abstract Recently, game theory has been shown to be useful for reasoning about real-world security set- tings where security forces must protect critical assets from potential adversaries. In fact, there have been a number of deployed real-world applications of game theory for security (e.g., AR- MOR at Los Angeles International Airport and IRIS for the Federal Air Marshals Service). Here, the objective is for the security force to utilize its limited resources to best defend their critical assets. An important factor in these real-world security settings is that the adversaries involved are humans who may not behave according to the standard assumptions of game-theoretic models. There are two key shortcomings of the approaches currently employed in these recent applica- tions. First, human adversaries may not make the predicted rational decision. In such situations, where the security force has optimized against a perfectly rational opponent, a deviation by the human adversary can lead to adverse aects on the security force’s predicted outcome. Second, human adversaries are naturally creative and security domains are highly dynamic, making enu- meration of all potential threats a practically impossible task and solving the resulting game, with current leading approaches, would be intractable. My thesis contributes to a very new area that combines algorithmic and experimental game- theory. Indeed, it examines a critical problem in applying game-theoretic techniques to situations ix where perfectly rational solvers must address human adversaries. In doing so it advances the study and reach of game theory to domains where software agents and humans may interact. More specifically, to address the first shortcoming, my thesis presents two separate algorithms to address potential deviations from the predicted rational decision by human adversaries. Ex- perimental results, from a simulation that is motivated by a real-world security domain at Los Angeles International airport, demonstrated that both of my approaches outperform the currently deployed optimal algorithms which utilize standard game-theoretic assumptions and additional alternative algorithms against humans. In fact, one of my approaches is currently under evalua- tion in a real-world application to aid in resource allocation decisions for the United States Coast Guard. Towards addressing the second shortcoming of enumeration of a large number of potential adversary threat capabilities, I introduce a new game-theoretic model for eciency, which addi- tionally generalizes the previously accepted model for security domains. This new game-theoretic model for addressing human threat capabilities has seen real-world deployment and is under eval- uation to aid the United States Transportation Security Administration in their resource allocation challenges. x Chapter 1: Introduction Security agencies are tasked with the important challenge of protecting critical infrastructure nationwide. For instance, security agencies are required to protect transportation networks, in- cluding the passengers, from potential terrorist activities or other disruptive activities that may halt the transportation of goods and people[DHS, 2012b]. Security agencies also protect critical environmental assets. For example, agencies may be responsible for protecting national parks and wildlife from illegal animal poaching or forest extraction[GTI, 2012; EPA, 2012]. In urban neighborhoods these agencies are responsible for protecting citizens by preventing criminal ac- tivity[LAPD, 2012]. Additionally, security agencies are responsible for stopping illegal flows of drugs, weapons, and more from entering or exiting national borders along with protecting legal assets from being seized through piracy or other means[DHS, 2012a]. While there are many more important security scenarios, the common problem among them is that security agencies only have limited resources to provide this critical protection. 1.1 Problem Addressed Stackelberg games, which were first introduced to model leadership and commitment [von Stack- elberg, 1934], have become popular as an approach to address critical security scenarios similar 1 to those previously mentioned. Recently, a game-theoretic approach based on Stackelberg games has been successfully used to model the security challenge of optimally allocating limited secu- rity resources across a set of potential targets [Paruchuri et al., 2008; Conitzer and Sandholm, 2006; Basilico et al., 2009]. While I provide a formal definition for Stackelberg games in Section 2.1, they are a natural model for such security problems because they model the commitment a defender (e.g., security agency) must make in allocating her resources before an attacker can conduct surveillance and choose his best attack strategy, considering the action chosen by the defender 1 . Indeed, this approach has been featured in multiple deployed real-world applications. Two prominent ones include a deployed application for randomizing checkpoints and canine patrols at the Los Angeles International airport since August of 2007 known as ARMOR and a deployed application for randomizing Federal Air Marshals on flights since 2009 known as IRIS [Jain et al., 2010]. This approach is also being investigated for numerous other applications such as GUARDS for randomizing security activities at airports for the Transportation Security Admin- istration [Pita et al., 2011], TRUSTS for randomizing urban security in transit systems Yin et al. [2012b], and PROTECT for randomizing port security for the United States Coast Guard [Shieh et al., 2012]. Beyond the deployed real-world applications, Stackelberg games have been used to study security problems ranging from “police and robbers” scenarios [Gatti, 2008], to com- puter network security [Lye and Wing, 2005], to missile defense systems [Brown et al., 2005] and terrorism [Sandler and M., 2003]. The development of these applications has inspired theoretical and algorithmic progress in ef- ficiently applying a game-theoretic framework based on Stackelberg games to real-world security 1 By convention, I refer to the defender (leader) as she and attacker (follower) as he 2 scenarios. While they eciently model strategic interactions between a defender and potential attacker, one possible failing of current models used in these deployed security applications [Jain et al., 2010; Yin et al., 2012b; Pita et al., 2011] is that they make strict assumptions on the un- derlying games 2 . In real-world domains, the strict assumptions made in the underlying models may lead to significant degradation in performance. In my work I examine three of the strict assumptions made in standard game-theoretic models against human adversaries. Namely, that the human adversary is perfectly rational, that they perfectly observe the leader’s strategy, and that their action space is tractable. Rationality Assumption: One of the strict assumptions made by standard game-theoretic approaches based on Stackelberg games is that the adversary is perfectly rational[Jain et al., 2010; Paruchuri et al., 2008; Conitzer and Sandholm, 2006]. That is, the adversary maxi- mizes his expected value based on the information available. However, human adversaries may not be expected-value-maximizers, computing optimal decisions. Instead, their deci- sions may be governed by their bounded rationality [Simon, 1956], which causes them to deviate from their expected optimal strategy. Such unexpected deviations can have arbi- trarily negative impacts on the security force’s expected value if not specifically accounted for. Observability Assumption: In standard Stackelberg games it is assumed that the adversary perfectly observes the security force’s strategy. In real-world settings, humans may have limited observability of the security force’s strategy, giving them a false impression of that 2 One exception is the application deployed since 2012 for the United States Coast Guard [Shieh et al., 2012], which relaxes these strict assumptions by utilizing an algorithm, MATCH, that I will present in this thesis to randomize patrols. 3 strategy. Furthermore, given limited information, humans can be biased in their decision making causing them to deviate from the optimal response of the observed strategy. Action Space Assumption: Game-theoretic models require defined (i.e., fixed) potential actions on behalf of all the players involved in order to compute an optimal mixed strategy for the security force. However, real-world security settings are dynamic in nature and humans are highly adaptable and creative allowing for the action space of an attacker to constantly be evolving. Furthermore, even if the action space was held static, the sheer number of potential attack methods by human adversaries may be too large to enumerate and solve in practice. Thus, intelligent methods for compactly representing the potential threat capabilities of human adversaries are required. In summary, given the standard strict assumptions, humans may deviate from the predicted optimal response, which can lead to significant losses on behalf of the security force. These losses are two-fold since the security force has not optimized against this sub-optimal play, but also since the security force could have made significant gains by exploiting potential human biases. In addition, enumerating all their threat capabilities may be intractable making a solution, in general, practically infeasible to compute. Accounting for these potential deviations and large action capabilities can provide significant benefits to a security force by providing solutions that help prevent large potential losses from unpredicted behavior and exploit human biases. 1.2 Contributions To address these three fundamental assumptions in dealing with humans my thesis contributes to a novel area of research that combines algorithmic and experimental game theory. My work 4 focuses on the very new and growing real-world problem of designing game-theoretic approaches to address human adversaries. Such human adversaries may not behave according to the strict assumptions made in standard game-theoretic models and thus may deviate from the predicted behavior or create complex problems that cannot be solved eectively. Specifically, I have made contributions in two key areas. The first key area of contribution is in developing algorithms to address potential deviations by human adversaries due to their bounded rationality or limited observability. These algorithms help address the first two strict assumptions presented and are presented as COBRA and MATCH. The second key area of contribution is in developing an alternative concept for addressing a human’s exceedingly large and evolving action space within a game-theoretical model for eciency. This new method for representing such a massive number of threat capabilities on behalf of the human adversary helps address the third strict assumption and I refer to this model as Security Circumvention Games. 1.2.1 COBRA/MATCH My new algorithms aid in addressing two strict assumptions made by optimal Stackelberg solvers that an adversary is both perfectly rational and that he perfectly observes the defender strat- egy. These algorithms combine key ideas from: (i) previous best known algorithms for solving Bayesian Stackelberg games [Paruchuri et al., 2008], (ii) robustness approaches for games from robust optimization literature [Aghassi and Bertsimas, 2006; Ord´ o˜ nez and Stier-Moses, 2007], and (iii) anchoring theories on human perception of probability distributions from psychology [Fox and Clemen, 2005]. While the robustness approaches help to address human response im- precision, anchoring, which is an expansion of support theory [Tversky and Koehler, 1994], helps address limited observational capabilities. In addition to its critical algorithmic component, my 5 work also diers from other behavioral based approaches in that I only consider potential devi- ations on behalf of the adversary while computationally assisting the security force in making rational decisions. Other behavioral approaches consider potential deviations on behalf of all players involved [Camerer, 2003; Selten, 1988; McKelvey and Palfrey, 1995]. The ideas pre- sented in these algorithms are generalizable and in fact one of the robustness approaches I present has been used in the work of Yang et al. [2011] for an algorithm known as RPT. This thesis introduces a mixed-integer linear program (MILP) COBRA (Combined Observ- ability and Rationality Assumption), that builds on a previous Bayesian Stackelberg game solver known as DOBSS, which uses the standard strict assumptions [Paruchuri et al., 2008]. This MILP continues to handle adversary reward uncertainty by utilizing Bayesian Stackelberg games, how- ever, it also addresses the uncertainty that may arise from human imprecision in choosing the expected value maximizing strategy due to bounded rationality and limited observations. As a result of this bounded rationality, the adversary may select an-optimal response strategy, i.e., the adversary may choose any of the responses within-reward of his optimal strategy. This choice may be caused by a variety of reasons, but COBRA attempts to guard against the choices that fall within this-bound of the optimal response. More specifically, given multiple possible-optimal responses, the robust approach is to assume that the attacker could choose the one that provides the security force the worst reward – not necessarily because the adversary attends to the security force’s reward, but to robustly guard against the worst-case outcome. In an adversarial setting, handling the worst case outcome may be in the best interest of the security force. This worst-case assumption contrasts with those of other Stackelberg solvers which assume the adversary will play a strong Stackelberg equilibrium (choosing a strategy that favors the leader in the case of 6 a tie) [Conitzer and Sandholm, 2006; Paruchuri et al., 2008], making COBRA novel to address human adversaries. The COBRA algorithm additionally utilizes the idea of anchoring biases [Fox and Clemen, 2005; Fox and Rottenstreich, 2003; See et al., 2006b], which are based upon support theory [Tversky and Koehler, 1994], to protect against limited observation conditions when addressing human adversaries. An anchoring bias is when, given no information about the occurrence of a discrete set of events, humans will tend to assign an equal weight to the occurrence of each event (a uniform distribution). This is also referred to as giving full support to the ignorance prior [Fox and Rottenstreich, 2003]. It has been shown through extensive experimentation that humans are particularly susceptible to giving full support to the ignorance prior before they are given any information and that, once given information, they are slow to update away from this assumption [Fox and Rottenstreich, 2003]. Thus, COBRA computes a mixed strategy for the defender considering how the human will perceive this strategy based on his bias toward the ignorance prior. While the COBRA algorithm demonstrates how accounting for potential deviations can im- prove the performance of defenders against human adversaries, I also present the MATCH al- gorithm which evolves the robustness approach of COBRA. Similar to COBRA, the MATCH algorithm avoids the complex task of explicitly modeling human decision making, which can be a dicult task in the absence of substantial human-decision-making data. MATCH leverages and modifies standard robust optimization techniques [Aghassi and Bertsimas, 2006] to create a new type of graduated optimization. Here, the defender’s loss for a potential deviation by the adversary is bounded by the distance of that deviation from the expected-value-maximizing strat- egy. That is, instead of attempting to predict exactly how a human decision maker will deviate, 7 MATCH limits the degradation of the defender’s reward based on how far the human adversary deviates. In doing so, MATCH addresses the potential loss to the defender for unpredicted de- viations, but still attempts to optimize against rational play. In the MATCH formulation we can control how much the degradation is limited for deviations providing a trade o between ro- bustness to deviations and the optimal expected value of the defender for rational play by the adversary. It is important to note that, unlike COBRA, MATCH is not designed to address obser- vational uncertainty and it is left for future work to investigate whether adding human anchoring biases into the MATCH formulation would outperform COBRA under limited observation. In order to evaluate the usefulness of these algorithms against human adversaries, I conducted extensive experiments with human subjects playing a security game motivated by a real-world security scenario at Los Angeles International airport (LAX). The human subjects played as an attacker who had to choose 1 of 8 possible gates to attack that were guarded by three guards. The experimental results demonstrate that both COBRA and MATCH outperform optimal algo- rithms based on the standard game-theoretic assumptions. Furthermore, my analysis reveals that MATCH is able to outperform the current best performing algorithm for addressing human ad- versaries under no observational uncertainty[Yang et al., 2011], establishing the benefits of such an approach when sucient models of human decision making are not available. 1.2.2 Security Circumvention Games In this thesis I present an alternative concept for addressing a human’s exceedingly large and evolving action space within a game-theoretic model for eciency. I developed a model for security settings called Security Circumvention Games (SCGs) that utilizes this new concept for defining attacker actions. Before the development of SCGs, all previously deployed work [Jain 8 et al., 2010] was based on a game-theoretic model known as security games [Korzhyk et al., 2011], which only allow for a single security activity and a single threat scenario. While it was ecient for the domains considered, it is not realistic in many other real-world domains. SCGs extend security games in a significant way, making the following contributions: (i) the ability to reason over heterogeneous security activities for a single target, and (ii) the ability to reason about an intelligent attacker who chooses threats designed to circumvent specific security activities. By allowing attackers circumvention capabilities we no longer need to consider specific threats, but rather which security measures they will attempt to avoid in executing a threat. By representing threats as circumvention strategies, SCGs are able to represent large-scale games in a compact way that can be solved by current Stackelberg solution techniquesParuchuri et al. [2008]; Pita et al. [2011]. In order to evaluate the value of the SCG model I examined two important issues. The first issue is with scalability and runtimes. In order for my new model to be useful it needed to be able to solve problems on the scale of real-world problems. The second issue is evaluating the value of the security policies generated using the assumptions in SCGs against potential alternative approaches. Thus, I examine SCGs against some potential standard alternative approaches for the security domains SCGs are designed for. By addressing these fundamental challenges it is possible to improve resource allocation in real-world security settings against human adversaries. In fact, my research has been applied in the real-world. Indeed, SCGs form the core of GUARDS, a scheduling assistant under evaluation by the United States Transportation Security Administration (TSA) since Fall 2010 to schedule limited security resources within airports nationwide [Pita et al., 2011]. Furthermore, MATCH 9 has been deployed in the PROTECT [Shieh et al., 2012] application to assist the United States Coast Guard with randomization of their patrolling routes since Spring 2012. 1.3 Guide to Thesis This thesis is organized in the following way. Chapter 2 introduces necessary background for the research presented in this thesis and Chapter 3 presents related work. Chapter 4 presents the algorithm COBRA and corresponding experimental results. Chapter 5 presents the algorithm MATCH and corresponding experimental results. Chapter 6 examines Security Circumvention Games, describing the model framework and justification, and compact representations for e- ciency. Finally, Chapter 7 concludes the thesis and presents issues for future work. 10 Chapter 2: Background The work in this thesis builds on Stackelberg games for modeling security domains. In Section 2.1 I will define a Stackelberg game for security domains and in Section 2.2 I will generalize it to a Bayesian Stackelberg game for security domains. In Section 2.3 I will describe the standard solution concept known as a strong Stackelberg equilibrium and in Section 2.4 I will introduce DOBSS, a general Stackelberg solver [Paruchuri et al., 2008] and the foundation for my COBRA and MATCH algorithms, along with two baseline algorithms (UNIFORM and MAXIMIN) for experimental comparison, which could be used in a Stackelberg setting. In Section 2.5 I will introduce the leading competitor to my approaches known as BRQR. In Section 2.6 I will define a constrained version of Stackelberg games known as security games. Finally in Section 2.7 I will describe a real-world security problem at Los Angeles International Airport (LAX) that motivates the experimental setup in this thesis and in Section 2.8 I will justify our choice of human subjects for these experiments. 2.1 Stakelberg Games In a Stackelberg game, a leader commits to a strategy first, and then followers sequentially self- ishly optimize their rewards, considering the action chosen by the leader. For the remainder of 11 this thesis I will refer to the leader as ’her’ and the follower as ’him’. To see the advantage of being the leader in a Stackelberg game, consider a simple game with the payo table as shown in Table 2.1, which was first presented by Conitzer and Sandholm [2006]. The leader is the row player and the follower is the column player. The only pure-strategy Nash equilibrium for this game is when the leader plays a and the follower plays c, which gives the leader a payo of 2; in fact, for the leader, playing b is strictly dominated. However, if the leader can commit to playing b before the follower chooses his strategy, then the leader will obtain a payo of 3, since the follower would then play d to ensure a higher payo for himself. If the leader commits to a uniform mixed strategy of playing a and b with equal (0.5) probability, then the follower will play d, leading to a payo for the leader of 3.5. c d a 2,1 4,0 b 1,0 3,2 Table 2.1: Payo table for example Stackelberg game. The Stackelberg games I consider in this thesis have two agents, the leader (defender), , and the follower (attacker/adversary), . Each player has a set of possible pure strategies, denoted =f 1 ;:::; j j g for the leader and =f 1 ;:::; j j g for the follower. A mixed strategy allows a player to play a probability distribution over pure strategies, denoted x2 X for the leader and q2 Q for the follower. Here, x i 2 [0; 1] represents the probability with which the leader takes pure strategy i 2 and similarly q i 2 [0; 1] represents the probability with which the follower takes pure strategy i 2 . I denote by S the index set over the leader’s pure strategies and 12 by S the index set over the follower’s pure strategies. Payos are respectively defined for the leader and the follower over all possible joint pure strategy outcomes: : !R (2.1) : !R (2.2) The payo functions are extended to mixed strategies in the standard way, by taking the expecta- tion over pure-strategy outcomes. That is, given a leader strategy x2 X and a follower strategy q2 Q the leader’s and follower’s payos can be determined by: (x; q) = X i2S X j2S ( i ; j ) x i q j (2.3) (x; q) = X i2S X j2S ( i ; j ) x i q j (2.4) 2.2 Bayesian Stackelberg Games The Bayesian extension to Stackelberg games allows for each player (leader or follower) to be one of multiple possible types, with each type associated with its own payo values. In this thesis the defender (leader), , only has one type since she is considering her own personal resources. However, the attacker (follower) can be one of a set of possible types denoted by 2 . For example, a security force may be interested in protecting against potential terrorist attacks and catching potential drug smugglers, which represent two dierent types of adversaries. Each type is represented by a dierent and possibly uncorrelated payo matrix for both the leader and follower. That is, the leader’s payos will vary along with the follower’s payos for each type of 13 follower. At any time the leader does not know what follower type she will face, however, she is aware of the probability distribution over follower types (i.e., she knows how frequently she will face each follower type). The probability with which follower type 2 appears is denoted by p . The follower is always aware of his own type and thus always has perfect information about the leader’s payos and his own payos. While the set of possible pure strategies remains the same for both the leader and the follower ( and respectively), payos for each player are now defined over all possible pure strategy outcomes for each follower type: : !R (2.5) : !R (2.6) Additionally, each follower type 2 can choose his own mixed strategy denoted by q 2 Q. For a particular follower type 2 , the payo functions are still extended to mixed strategies in the standard way, by taking the expectation over pure-strategy outcomes. That is, given a follower type 2 , a leader strategy x2 X and a follower strategy q 2 Q the leader’s and follower’s payos can be determined by: (x; q ; ) = X i2S X j2S ( i ; j ; ) x i q j (2.7) (x; q ; ) = X i2S X j2S ( i ; j ; ) x i q j (2.8) 14 Given this formal model, the leader’s goal is to determine the mixed strategy x2 X, such that her expected value is maximized given that each follower type will choose his expected-value- maximizing action with complete knowledge of the leader’s mixed strategy. Thus, the follower’s strategy in a Stackelberg game is a function that selects a strategy in response to each leader strategy for each follower type: F : X ! Q (2.9) Such a commitment to a mixed strategy models a real-world situation where security forces com- mit to a randomized patrolling strategy first. Given this commitment, an adversary can conduct as much surveillance of this mixed strategy as he desires. Even with knowledge of this mixed strat- egy, the adversary has no specific knowledge of what the security force may do on a particular day however. He only has knowledge of the mixed strategy the security force will use to decide her resource allocations for that day. In this model, predictable defense strategies are vulnerable to exploitation by a determined adversary. 2.3 Strong Stackelberg Equilibrium Leitmann [1978] proposed two types of unique Stackelberg equilibria, which are typically called “strong” and “weak” after Breton et al. [1988]. The strong form assumes that the follower will always choose the optimal strategy for the leader in cases of indierence, while the weak form assumes that the follower will choose the worst strategy for the leader in cases of indierence. A strong Stackelberg equilibrium exists in all Stackelberg games, but a weak Stackelberg equilib- rium may not [Basar and Olsder, 1995]. In addition, the leader can often induce the favorable 15 strong equilibrium by selecting a strategy arbitrarily close to the equilibrium that causes the follower to strictly prefer the desired strategy [von Stengel and Zamir, 2004]. Recent research utilizing a Stackelberg framework with the standard assumptions of rationality and perfect ob- servability for security domains has adopted strong Stackelberg equilibrium [Kiekintveld et al., 2009; Jain et al., 2011b]. Definition 1. A pair of strategies (x; F ) form a Strong Stackelberg Equilibrium if they satisfy the following: 1. The leader plays a best response: P 2 (x; F (x; ); ) P 2 (x 0 ; F (x 0 ; ); ) 8 x 0 2 X 2. The follower plays a best-response: P 2 (x; F (x; ); ) P 2 (x; q; ) 8 x2 X; q2 Q 3. The follower breaks ties optimally for the leader: P 2 (x; F (x; ); ) P 2 (x; q; ) 8 x2 X; q2 Q (x), where Q (x) is the set of follower best-responses to x, as above. 2.4 DOBSS and Baseline Algorithms For completeness this thesis includes both a uniformly random strategy and a MAXIMIN strategy against human opponents as a baseline against the performance of both existing algorithms, such as DOBSS, and the new algorithms presented. Proposed algorithms must outperform the two baseline algorithms to provide benefits. 16 2.4.1 DOBSS I now describe the Decomposed Optimal Bayesian Stackelberg Solver (DOBSS) in detail as it provides a starting point for COBRA. While the problem of choosing an optimal strategy for the leader in a Stackelberg game is NP-hard for a Bayesian game with multiple follower types [Conitzer and Sandholm, 2006], researchers have continued to provide practical improvements. DOBSS is an ecient general Stackelberg solver [Paruchuri et al., 2008] and is in use for security scheduling at the Los Angeles International Airport. It operates directly on the compact Bayesian representation, giving speedups over the multiple linear programs method [Conitzer and Sand- holm, 2006], which requires conversion of the Bayesian game into a normal-form game by the Harsanyi transformation [Harsanyi and Selten, 1972]. In particular, DOBSS obtains a decompo- sition scheme by exploiting the property that follower types are independent of each other. The key to the DOBSS decomposition is the observation that evaluating the leader strategy against a Harsanyi-transformed game matrix is equivalent to evaluating against each of the game matrices for the individual follower types and then obtaining a weighted sum. I first present DOBSS in its most intuitive form as a Mixed-Integer Quadratic Program (MIQP); I then present a linearized equivalent Mixed-Integer Linear Program (MILP). The DOBSS model explicitly represents the actions by the leader and the optimal actions for the follower types in the problem solved by the leader. Note that DOBSS needs to consider only the expected-value- maximizing pure strategies of the follower types, since for a given fixed mixed strategy x2 X of the leader, each follower type, 2 , faces a problem with fixed linear rewards. If a mixed strategy is optimal for the follower, then so are all the pure strategies in support of that mixed strategy. 17 In the MIQP presented below, let M be a large positive number. Given prior probabilities p , with 2 , of facing each follower type, the leader solves the following: max X i2S X j2S X 2 p ( i ; j ; ) x i q j s:t: X i2S x i = 1 (2.10) x i 2 [0::: 1] 8 i2 S (2.11) X j2S q j = 1 8 2 (2.12) q j 2f0; 1g 8 2 ; j2 S (2.13) 0 a X i2S ( i ; j ; ) x i (1 q j ) M 8 2 ; j2 S (2.14) a 2R 8 2 (2.15) Here for a leader strategy x and a strategy q for each follower type, the objective represents the expected reward for the leader considering the a priori distribution over dierent follower types p . Constraints 2.10 and 2.11 define the set of feasible solutions x2 X as a probability distribution over the set of pure strategies i 2 . Constraints 2.12 and 2.13 limit the mixed strategies of follower type , q 2 Q, to be only pure strategies over the set (that is each q has exactly one coordinate equal to one and the rest equal to zero). The two inequalities in Constraint 2.14 ensure that q j = 1 only for a strategy j that is optimal (i.e, expected-value-maximizing) for follower type . Indeed this is a linearized form of the optimality conditions for the linear programming problem solved by each follower type. Constraint 2.14 can be explained as follows: 18 the leftmost inequality ensures that8 2 ; j2 S : a P i2S ( i ; j ; )x i . This means that given the leader’s strategy x, a is an upper bound on follower type ’s reward for any strategy. The rightmost inequality is inactive for every strategy where q j = 0, since M is a large positive quantity. For the strategy that has q j = 1 this inequality states a P i2S ( i ; j ; ) x i , which combined with the previous inequality shows that this strategy must be optimal for follower type . This quadratic programming problem can be linearized through the change of variables z i j = x i q j , thus obtaining the mixed integer linear programming problem presented below. My imple- mentation of DOBSS solves this MILP, which was shown to be equivalent to the quadratic pro- gramming problem presented above and the equivalent Harsanyi transformed Stackelberg game by Paruchuri et al. [2008]. The DOBSS MILP can be solved using ecient integer programming packages. For a more in depth explanation of DOBSS please see Paruchuri et al. [2008]. While DOBSS is an ecient general Stackelberg solver, it has three fundamental problems in the mixed strategies it computes given a human adversary (follower). First, DOBSS assumes the human adversary will break ties in the defender’s (leader’s) favor. However, in real-world security scenarios, which are adversarial in nature, it is plausible that a human adversary would not make any dierentiation between ties or possibly even break ties against the defender leading to a reduced defender reward. Second, DOBSS assumes the human adversary is perfectly rational and thus will choose an expected-value-maximizing strategy among all of his alternatives. However, if the human adversary is boundedly rational and deviates from an expected-value-maximizing strategy it can severely impact the defender’s expected value. Finally, in real-world security scenarios it may be dicult for a human adversary to perfectly observe the defender’s strategy. 19 Such imperfect observation can lead the human adversary to deviate from the expected-value- maximizing strategy, which can again have adverse aects on the defender’s expected value. This thesis helps address these critical limitations in the DOBSS formulation. max X i2S X j2S X 2 p ( i ; j ; ) z i j s:t: X i2S X j2S z i j = 1 8 2 (2.16) X j2x z i j 1 8 2 ; i2 S (2.17) z i j 2 [0::: 1] 8 2 ; i2 S ; j2 S (2.18) q j X i2S z i j 1 8 2 ; j2 S (2.19) X j2S q j = 1 8 2 (2.20) q j 2f0; 1g 8 2 ; j2 S (2.21) 0 a X i2S ( i ; j ; )( X h2S z ih ) (1 q j ) M 8 2 ; j2 S (2.22) X j2S z i j = X j2S z 1 i j 8 2 ; i2 S (2.23) a 2R 8 2 (2.24) 2.4.2 UNIFORM UNIFORM is the most basic method of randomization which just assigns an equal probability of taking each strategy i 2 (a uniform distribution). 20 2.4.3 MAXIMIN MAXIMIN is a traditional approach which assumes the follower may take any of the available actions. The objective of the following linear program is to maximize the minimum reward, V, the leader will obtain irrespective of the follower’s action: max X 2 p V s:t: X i2S x i = 1 (2.25) X i2S ( i ; j ; ) x i V 8 2 ; j2 S (2.26) x i 2 [0::: 1] 8 i2 S (2.27) 2.5 BRQR Best Response to Quantal Response (BRQR) is based on the quantal response (QR) model [McK- elvey and Palfrey, 1995] of human decision making. This QR model is a well-founded solution concept in game theory derived from Nobel-prize-winning work in choice modeling theory [No- belprize.org, 2012], and there has been significant support for this QR model elsewhere in the literature [Wright and Leyton-Brown, 2010]. The QR model suggests that instead of strictly maximizing utility, individuals respond stochastically in games: the chance of selecting a non- optimal strategy increases as the cost of such an error decreases. Specifically, the QR model defines a parameter,, which represents the amount of noise in a player’s response. 21 In applying the QR model, BRQR only adds noise to the response function for the adversary, so the defender computes an optimal strategy assuming the attacker responds with a noisy best- response. Given and the defender’s mixed-strategy x, the adversaries’ quantal response q j (i.e., the probability of taking pure strategy j 2 ) can be written as: q j = e P i2S ( i ; j )x i P k2S e P i2S ( i ; k )x i (2.28) The goal is to maximize the defender’s expected utility given q (i.e., P i2S P j2S ( i ; j ) x i q j ). Combined with Equation 2.28, the problem of finding the optimal mixed strategy for the defender can be formulated as: max X i2S X j2S ( i ; j ) x i e P i2S ( i ; j )x i P k2S e P i2S ( i ; k )x i s:t: X i2S x i = 1 (2.29) x i 2 [0::: 1] 8 i2 S (2.30) A major diculty of this modeling approach however, is that it requires the appropriate es- timation of to determine the level of noise in the human adversary’s response function. Yang et al. [2011] have proposed a method for appropriately estimating within security settings. Still, in real-world scenarios that are often complex and large in scope, creating an accurate model of human decision making or estimating such parameters can be a dicult task. This diculty is exacerbated in security settings where information about potential adversaries is often sparse and noisy. 22 2.6 Security Stackelberg Games At this time I introduce a restricted version of a Stackelberg game known as a security game [Korzhyk et al., 2011]. The MATCH algorithm that I will present in Chapter 5 is designed specif- ically for this restricted class of games because in a general Stackelberg framework a solution to the MATCH algorithm may not exist. Security games make strict assumptions on the payos of both the leader and the follower in a Stackelberg game with security settings specifically in mind. In security games, there are a set of targets T =ft 1 ;:::; t jTj g that the defender wants to protect from an attacker and a set of resources R =fr 1 ;:::; r jRj g (e.g., police ocers) that the defender may deploy to protect the targets. Resources are identical in that any resource can be deployed to protect any target, and provide equivalent protection to that target. Furthermore, there is no additional benefit to deploying more than one resource to any individual target (i.e., once a target is protected by a resource it is considered fully protected). The payos for the defender and attacker depend on which target is attacked, and whether the target is protected (covered) or not (uncovered). Formally, the defender’s and attacker’s payos are defined for each target depending on whether it is covered (c) or uncovered (u): U u : T!R (2.31) U c : T!R (2.32) U u : T!R (2.33) U c : T!R (2.34) 23 Let Equation 2.35 denote the dierence between the defender’s covered and uncovered payos. Similarly, Equation 2.36 denotes the dierence between the attacker’s uncovered and covered payos. As a key property of security games, it is necessary that U (t i ) > 0 and U (t i ) > 0. In other words, adding resources to cover a target helps the defender and hurts the attacker. U (t i ) = U c (t i ) U u (t i ) 8 t i 2 T (2.35) U (t i ) = U u (t i ) U c (t i ) 8 t i 2 T (2.36) For the purpose of this thesis I only consider security games that have schedules of size 1 (i.e., satisfy the Subsets of Schedules Are Schedules property [Korzhyk et al., 2011]). A pure strategy of the defender then, still denoted by i 2 , is a subset of targets from T with size equal tojRj. An adversary’s pure strategy, still denoted by j 2 , is exactly one target t j 2 T. Given that I only consider security games with schedules of size 1, Kiekintveld et al. [2009] have shown that the defender’s possible mixed strategies can be transformed into an equivalent coverage vectors of probability, c2 C, where c i is the marginal probability of covering t i and similarly a coverage vector can be transformed into an equivalent valid mixed strategy x2 X of the defender[Korzhyk et al., 2011]. Given that payos for both the defender and attacker only depend on whether a particular target t i 2 T is covered or not, the marginal probability c2 C can be used in place of the mixed strategy x2 X to compute the defender’s optimal resource allocation, reducing the complexity of the defender’s actions. Specifically, there are jTj jRj potential pure strategies and onlyjTj marginal probabilities. Where before P i2S x i = 1 for a defender’s mixed strategy, it is now the case that P i2T c i =jRj and c i 2 [0::: 1] 8 i2 T for the defender’s coverage vector. In other words, the 24 defender wants to use all of her resources and placing more than 1 resource on a target (i.e., more that 100% coverage) is not beneficial. Here, given a coverage vector c2 C and an attacker mixed strategy q 2 Q, the defender’s and attacker’s payos are defined in Equations 2.37 and 2.38 respectively. Additionally, as Kiekintveld et al. [2009] have previously defined, the attack set, (c), contains all targets that yield the maximum expected value for the attacker given coverage c as shown in Equation 2.39. In a SSE, the attacker selects the target in the attack set with the maximum expected value for the defender, which I denote by t . It follows that the expected SSE value for the defender is ˆ U (c) = U (c; t ) and for the attacker ˆ U (c) = U (c; t ). U (c; q) = X i2S q i (c i U c (t i ) + (1 c i )U u (t i )) (2.37) U (c; q) = X i2S q i (c i U c (t i ) + (1 c i )U u (t i )) (2.38) (c) =ft : U (c; t) U (c; t 0 ) 8 t 0 2 Tg (2.39) While COBRA builds o the general Stackelberg solver DOBSS, MATCH leverages the unique structure of security games to provide improved runtime benefits in computing the de- fender’s resource allocation. MATCH is able to reason over the marginal coverage vector, greatly reducing the defender’s action space from jTj jRj potential actions tojTj potential actions. In doing so, MATCH is able to solve for larger scale problems, which is necessary in the deployed PRO- TECT application for the US Coast Guard [Shieh et al., 2012] that uses the MATCH algorithm. 25 2.7 Los Angeles International Airport While there are a number of applications that utilize a Stackelberg framework [Jain et al., 2010; Shieh et al., 2012; Pita et al., 2011], I chose to model my experimental setup after the security scenario at Los Angeles International Airport because it is the least constrained and simplest real-world deployed system making it ideal for an initial investigation against human subjects. I describe the specific challenges in the security problems faced by the Los Angeles World Airport (LAWA) police. Los Angeles International Airport (LAX) is the fifth busiest airport in the United States, the largest destination airport in the United States, and serves 60-70 million passengers per year [LAWA, 2012; Stevens et al., 2006]. LAX is unfortunately also suspected to be a prime terrorist target on the west coast of the United States, with multiple arrests of plotters attempting to attack LAX [Stevens et al., 2006]. To protect LAX, LAWA police have designed a security system that utilizes multiple rings of protection. As is evident to anyone traveling through LAX airport, these rings include such things as vehicular checkpoints, police units patrolling the roads to the terminals and inside the terminals (with canines), and security screening and bag checks for passengers. There are unfortunately not enough resources (police ocers) to monitor ev- ery single event at the airport; given its size and number of passengers served, such a level of screening would require considerably more personnel and cause greater delays to travelers. Thus, assuming that all checkpoints and terminals are not being monitored at all times, setting up avail- able checkpoints, canine units or other patrols on deterministic schedules allows adversaries to learn the schedules and plot an attack that avoids the police checkpoints and patrols, which makes deterministic schedules ineective. 26 (a) LAX Checkpoint (b) Canine Patrol Figure 2.1: LAX Security Randomization oers a solution here. In particular, from among all the security measures that randomization could be applied to, LAWA police have so far posed two crucial problems for the deployed ARMOR system [Pita et al., 2008]. First, given that there are many roads leading into LAX, where and when they should set up checkpoints to check cars driving into LAX. For example, Figure 2.1(a) shows a vehicular checkpoint set up on a road inbound towards LAX. Police ocers examine cars that drive by, and if any car appears suspicious, they do a more detailed inspection of that car. LAWA police wished to obtain a randomized schedule for such checkpoints for a particular time frame. For example, if they are to set up two checkpoints, and the time frame of interest is 8 AM to 11 AM, then a candidate schedule may suggest to the police that on Monday, checkpoints should be placed on route 1 and route 2, whereas on Tuesday during the same time slot, they should be on route 1 and 3, and so on. Second, LAWA police wished to obtain an assignment of canines to patrol routes through the terminals inside LAX. For example, if there are three canine units available, a possible assignment may be to place canines on terminals 1, 3, and 6 on the first day, but terminal 2, 4, and 6 on another day and so on based on the available information. Figure 2.1(b) illustrates a canine unit on patrol at LAX. 27 Given these problems, the ARMOR system considers three key challenges [Pita et al., 2008]: (i) potential attackers can observe security forces’ schedules over time and then choose their attack strategy – the fact that the adversary acts with knowledge of the security forces’ schedule makes deterministic schedules highly susceptible to attack; (ii) there is unknown and uncertain information regarding the types of adversary LAWA police may face; (iii) although randomization helps eliminate deterministic patters, it must also account for the dierent costs and benefits associated with particular targets. I use this real-world security problem and domain as the basis for my experimental setup. In particular I examine the canine patrolling problem of LAX where there are 8 terminals and 3 canines patrolling those terminals (although at LAX there may be more or less canines on patrol at any time). However, beyond the key challenges the ARMOR system addressed [Pita et al., 2008], I examine two critical assumptions of the underlying DOBSS algorithm in use for the ARMOR system [Pita et al., 2008]. Namely, that the human adversary is perfectly rational and perfectly observes the defender strategy. 2.8 Human Subjects Since my algorithms are centered on addressing non-optimal and uncertain human responses, traditional proofs of correctness and optimality are insucient: it is necessary to experimentally test these new approaches against existing approaches. Experimental analysis with human sub- jects allows me to show how my algorithms are expected to perform against human adversaries compared to alternative approaches. To that end, I perform my experiments against a general 28 population of undergraduate and graduate students in engineering at the University of South- ern California for COBRA and across a random demographic of people residing in the United States for MATCH. Given that the experimental results for COBRA and MATCH are across a random demographic of people, it is necessary to justify that these approaches are useful against terrorists specifically. This is particularly important since MATCH is already deployed for the United States Coast Guard. While it is virtually impossible to test my approaches against other approaches with actual terrorists who will behave according to their normal decision-making pro- cesses, I will argue that the experiments using students and United States residents provide results in the right direction. The key question is if the terrorist psychiatric profile may lead to vastly dierent decision- making processes than the general population. An argument could be made that terrorists are completely irrational or suer from severe psychosis causing them to make near arbitrary deci- sions. Thus they do not attempt to maximize expected-value or even attempt to make boundedly rational decisions making any game-theoretic approach, even one that considers potential de- viations due to bounded rationality, irrelevant. However, research has shown that the primary shared characteristic of the psychiatric profile of the decision makers of terrorist factions is their normalcy [Abrahms, 2008; Richardson, 2006; Gill and Young, 2011]. This extensive research suggests that terrorists are actually highly rational and carefully construct their attack strategies attempting to maximize their expected value, justifying a game-theoretic approach [Roso and John, 2009; Keeney and von Winterfeldt, 2010; Abrahms, 2008; Richardson, 2006]. As Louise Richardson, who has done significant research on the psyche of terrorists, states “But terrorists, by and large, are not insane at all. Their primary shared characteristic is their normalcy, insofar as we understand the term. Psychological studies of terrorists are virtually unanimous on this 29 point” [Richardson, 2006]. Furthermore, Abrahms [2008] both agrees on this point and notes that these studies are based on records of dozens of terrorist organizations from the late 1960s to the present. It follows that testing against the general population can give some strategic insights on how to better defend against potential terrorists attacks given the inability to test against actual terrorists. Based on the assumption that terrorists are rational agents, an argument could be made that it should be sucient to calculate a game-theoretic optimal resource allocation strategy. However, modern decision theory recognizes that human decision makers face cognitive and informational constraints. Additionally, terrorists sometimes face competing objectives and noisy information [Allison and Zelikow, 1999; Abrahms, 2008]. Thus, as my algorithms purpose, deviations are likely due to bounded rationality and cognitive limitations. While the algorithms presented in this thesis are designed with the real-world applications of Stackelberg games for preventing terrorist based activities specifically in mind, another benefit of experimenting with the general population is that the approaches presented in this thesis can be applied to a large number of criminal scenarios beyond terrorism. Criminal activities can be broadly broken down into six categories: (i) Property crimes, (ii) violent crimes, (iii) sex crimes, (iv) gangs and crime, (v) white-collar occupational crime, and (vi) drugs and crime [Pogrebin, 2012]. It is the responsibility of many dierent security agencies to attempt to prevent these criminal activities, with each agency focusing on dierent categories of crime. The approaches presented in this thesis can potentially aid a number of these agencies. For instance, urban police may need to patrol a large number of neighborhoods in an attempt to prevent property crimes. If the police take a deterministic approach, visiting each neighborhood in a set sequence, then a determined criminal could monitor their approach and exploit this pattern to potentially rob an 30 innocent victim. As another example, the Internal Revenue Service (IRS) may audit a company for fraudulent activities if they notice strange activity. However, the IRS could also use this approach to audit additional companies on a randomized basis in an attempt to catch any activity they may have missed. Given the large range of criminal based activities, the resulting types of human criminals will span a wide variety making the use of a general population largely applicable. In the future, this type of approach may be further refined for specific types of criminals or criminal activities. The idea is that a personality and demographic profile can be defined for crim- inals who participate in a particular type of criminal activity and we can experimentally evaluate the decision-making process for people of that profile. In doing so, we could tailor an approach to that particular decision-making process, thus improving the benefits. For example, research has shown that terrorists tend to be risk-averse and attempt to avoid uncertainty in planning their attack strategies [Morral and Jackson, 2009]. By studying risk-averse subjects it may be possible to improve upon the approaches presented in this thesis for the deployed applications discussed previously [Jain et al., 2010; Shieh et al., 2012]. However, whenever performing experimental evaluation of dierent approaches it can be particularly dicult to obtain the correct popula- tion that needs to be examined and this diculty has been considered before in other behavioral game-theoretic studies[Camerer, 2003]. 31 Chapter 3: Related Work The related work on game-theoretic approaches for security can be broadly divided into three categories: (i) ecient solutions for computing optimal Stackelberg equilibria, (ii) computing robust strategies, and (iii) addressing deviations from the theoretically optimal choice. 3.1 Computing Optimal Stackelberg Equilibria Numerous algorithms have been proposed for computing strong Stackelberg equilibria (SSE) strategies for Bayesian Stackelberg games. These can be further broken down into two categories: (i) algorithms for computing ecient solutions to general Bayesian Stackelberg games, and (ii) algorithms for computing ecient solutions for large-scale security games (large defender and attacker action spaces). This related work largely complements the work presented in this thesis since it is possible to combine the ecient approaches presented in this related work with my new approaches for addressing human adversaries. 3.1.1 Ecient Solutions to general Bayesian Stackelberg games Computing a solution to a Bayesian Stackelberg game has been previously shown to be an NP- hard problem, making scale-up a dicult challenge [Conitzer and Sandholm, 2006]. One of the 32 first optimal Bayesian Stackelberg solvers is the multiple-LPs approach, which solves many linear programs to compute the optimal defender strategy in Bayesian Stackelberg games [Conitzer and Sandholm, 2006]. Building on this approach, DOBSS [Paruchuri et al., 2008] decomposes the Bayesian game into individual types and solves a mixed-integer linear program to compute the strong Stackelberg equilibrium (SSE) strategy. In Section 2.4 I include an in depth examination of the DOBSS algorithm as it forms the foundation for the COBRA algorithm I will present. Fol- lowing in this direction, HBGS uses a novel technique based on hierarchical decomposition and branch and bound search over the follower type space and has been shown to be orders of mag- nitude faster than previous approaches [Jain et al., 2011b]. Similarly, Jain et al. [2011b] present a new exact algorithm called HBSA, which extends the previous fastest known security game solver towards the Bayesian case. While COBRA and MATCH do not utilize the approaches presented in HBGS and HBSA, it is possible that in the future these approaches could be applied to improve the runtime performance of both algorithms. The presented approaches address Bayesian Stackelberg games with a finite number of fol- lower types. However, work has also been done to address infinite or continuous Bayesian Stack- elberg games allowing for payos to be represented using continuous payo distributions as opposed to discrete payos. Kiekintveld et al. [2011] present several techniques for finding ap- proximate solutions for this class of games, and show empirically that these approaches demon- strate improvements over the alternative approaches presented. Kiekintveld et al. [2011] present SBE, SRD, GMC, and DTS as approaches for approximating solutions to such games and empir- ically evaluate the eectiveness of each against previous best known solvers. Building both upon the work of finite Bayesian Stackelberg games and infinite Bayesian Stackelberg games, Yin and Tambe [2012] present the HUNTER algorithm. The HUNTER algorithm was designed to both 33 address a discrete number of attacker types and, based on infinite Bayesian Stackelberg games, continuous uncertainty such as the follower’s observation noise, the leader’s execution error, and both players’ payo uncertainty. By considering observation noise, execution uncertainty, and payo uncertainty; HUNTER provides a type of robustness that is similar to COBRA and MATCH. However, there are some key dierences in the approach taken by HUNTER and that taken by COBRA and MATCH. HUNTER attempts to maximize the mean outcome of a known distribution of error or noise by using sample average approximation theories as opposed to maximizing the worst-case outcome. Furthermore, HUNTER still considers a perfectly rational adversary who will make the optimal choice given that error or noise in observation, execution, and/or payo. HUNTER does not consider that a human adversary may deviate even further due to bounded rationality or biases such as an anchoring bias given observation uncertainty. COBRA and MATCH consider specific types of deviations from human adversaries versus arbitrary deviations. Assuming there is no uncertainty, HUNTER’s approach may help in addressing human adver- saries who may deviate due to bounded rationality by artificially introducing some error. How- ever, this type of approach considers the mean of arbitrary deviations within that error level versus the directed deviations of COBRA and MATCH. In the future, combining the general robustness approach of HUNTER and the directed robustness of COBRA and MATCH may provide further benefits in domains where observation noise, execution uncertainty, and/or payo uncertainty are present even beyond the cognitive limitations of the human adversary. 34 3.1.2 Ecient Solutions for Large-Scale Security Games ORIGAMI and ERASER-C are algorithms that have been developed for larger and more complex security settings that exploit the underlying structure of security games specifically [Kiekintveld et al., 2009]. ORIGAMI is a polynomial time algorithm that computes optimal security force strategies for a security game when no scheduling constraints are present by exploiting the con- straints on the reward structure of the game. The ERASER-C mixed-integer linear program al- lows for certain kinds of resource and scheduling constraints. Formally, the set of legal schedules S =fs 1 ::: s l g is a subset of the power set of the targets, with restrictions on this set representing scheduling constraints. ERASER-C provides scale-ups by computing the security force coverage per schedule, instead of computing mixed strategies over a joint assignment for all security force resources. Unfortunately, Kiekintveld et al. [2009] show that ERASER-C may fail to generate a correct solution in cases where the security force may face arbitrary scheduling constraints. Improving upon this work, Jain et al. [2010] present ASPEN, which utilizes a novel branch- and-price approach to overcome both the scaling challenge and the challenge of generating cor- rect solutions given arbitrary scheduling constraints. Here, the algorithm starts by considering a minimal set of pure strategies for both the players (security force and adversary) and then itera- tively generate additional pure strategies that will improve the payo of the corresponding player (e.g., a security force’s pure strategy is added if it helps increase the security force’s payo). This process is repeated until the optimal solution is obtained. ASPEN has empirically been shown to both be competitive with ERASER-C for the restricted class of games where ERASER-C is applicable and to solve far more general instances of the scheduling problems where ERASER-C and other existing techniques fail. 35 Beyond the class of games where resources are assigned to static targets, there are also numer- ous patrolling problems and corresponding ecient solution techniques within the security game setting. Basilico et al. [2009] present an ecient game-theoretic approach utilizing a Stackelberg framework for optimal patrolling of arbitrary topologies. Similarly, Agmon et al. [2011] present an approach for patrolling a closed perimeter or open polyline (fence) to minimize an adversaries chance of penetrating the perimeter or open polyline. Johnson et al. [2012] examine a dier- ent patrolling problem where the goal is to maximize the radius of a successful patrol against would be forest extractors in order to preserve the largest pristine forest area. Similarly, Yin et al. [2012b] present an initial model and corresponding linear program for computing optimal patrol strategies in urban transit systems with certain temporal and spacial constraints. Finally, there exists a class of algorithms for scheduling resources on static targets within a graph based or network structure. Similar to the other work described, the goal here is to max- imize the security forces overall reward given that the adversary will attempt to find their way from some starting node to some target destinations with diering value. RANGER is a heuris- tic linear program for solving this complex problem, which eciently creates optimal marginal distributions for placing checkpoints [Tsai et al., 2010]. However, it is proven that the reward found by RANGER is an overestimate of the true optimal reward. Jain et al. [2011a] have shown that RANGER can be arbitrarily bad in general settings and have presented RUGGED, which is the first scalable optimal solution technique for such network security games. RUGGED, like ASPEN, utilizes a column generation approach where it starts by considering a minimal set of pure strategies for both players. 36 3.2 Computing Robust Strategies This line of research is most similar to that presented in this thesis, however, a crucial dierence is that this work considers robustness against a rational adversary in the face of certain types of uncertainty whereas I consider the case where the security force deals with a boundedly rational adversary who may deviate even without uncertainty in the game. Additionally, I consider the human adversary’s reaction to certain types of uncertainty. Korzhyk et al. [2011] consider the limiting case where an attacker has no observations and thus investigate the equivalence of Stackelberg vs Nash equilibria. Yin et al. [2012a] develop RECON to compute robust solutions in the case where there may be noise in the security force’s execution of the suggested mixed strategy and/or the observations made by an adversary can be noisy. RECON attempts to maximize the worse case outcome within some execution and/or ob- servation noise or error level from the intended strategy. Similarly, An et al. [2012] consider the condition when the adversary does not have perfect surveillance capabilities, and thus are un- able to learn the exact strategy of the defender. A new model is presented, security games with strategic surveillance (SGSS), which models the adversary’s belief update and strategic surveil- lance decisions in security games. In addition, An et al. [2012] provide multiple formulations for computing the defender’s optimal strategies, including non-convex programming and con- vex approximation, and provide an approximate approach for computing the adversary’s optimal surveillance length. Kiekintveld et al. [2011] model distributions over preferences of an adversary using infinite Bayesian games, and propose a number of algorithms (SBE, SRD, GMC, and DTS) to gener- ate approximate solutions for such games. A critical assumption of traditional game-theoretic 37 solvers is that both the security force and adversary are perfectly aware of all the payos in- volved. Here, the concept is to be robust to potential uncertainty in the adversary’s target payos. Building both upon the work for observation/execution uncertainty and payo uncertainty, Yin et al. [2012b] present the HUNTER algorithm. The HUNTER algorithm was designed to provide a unified method for creating robust strategies against the adversary’s observation noise, the secu- rity force’s execution error, and both players’ payo uncertainty. Again, here HUNTER attempts to maximize the mean outcome within a given level of uncertainty. Beyond execution, observation, and payo uncertainty, secrecy and deception have also been modeled for Stackelberg games [Zhuang and Bier, 2011]. Outside of Stackelberg games, models for execution uncertainty in game-theory have been separately developed [Archibald and Shoham, 2009]. Additionally, robust solution methods for simultaneous move games have been studied [Aghassi and Bertsimas, 2006; Porter et al., 2002]. 3.3 Addressing Suboptimal Decisions A common concern in game-theoretic settings has been considering potential deviations from the game-theoretic optimal play [Camerer, 2003; Selten, 1988; McKelvey and Palfrey, 1995]. One well known approach is trembling hand perfect equilibrium [Selten, 1988] where players find an equilibrium given that the other players may choose unintended strategies due to a “trembling hand” or error. A trembling hand perfect equilibrium is a sequence of perturbed games (games where only totally mixed strategies are allowed) that converge to some base game Nash equilibria. The idea here is that even if the second player trembles, it will not give the first player incentive 38 to deviate from the current Nash equilibrium and vice versa. Another approach is that of - equilibrium [Tijs, 1981] where it is assumed that in simultaneous move games an-equilibrium point has been reached when a unilateral deviation from that equilibrium point by one of the players will not increase the payo of that player by more than. In addition to these key equilibrium concepts, the field of experimental game theory has pro- vided a wealth of contributions and insights on deviations of human play from equilibrium pre- dictions and dierent models to explain these deviations [Camerer, 2003]. In contrast with these equilibrium concepts and human models, there are three key dierences in the work in my the- sis: (i) the equilibrium approaches seek stable equilibrium points where the players involved are not expected to deviate and thus do not specifically address the impact of unexpected deviations or protect against them. In contrast, my approaches robustly guard against potential deviations due to sub-optimal play; (ii) both the equilibrium concepts and human models consider strate- gies where all players involved may deviate whereas my approach computes a rational strategy for the defender (i.e., an approach that the defender is expected to follow without deviation) by utilizing insights on the potential deviations of a boundedly rational human opponent; and (iii) the equilibrium approaches only define stable equilibrium whereas my work also considers the computational aspects of finding an optimal strategy. More similar to the work presented in this thesis, there has been significant work in developing agents that consider interactions with humans[Haim et al., 2012; Azaria et al., 2012; de Melo et al., 2011]. Here, the agents’ goal is to improve the humans performance or decision making on some task based on the agents’ interaction with that human. For instance, Azaria et al. [2012] consider the design of an automated advice provision agent who will have repeated interactions with a human user. Similarly, de Melo et al. [2011] consider the eect of agents with emotional 39 expression when negotiating with humans. Marsella et al. [2010] consider the role and history of computational models of human emotion in general and their wide variety of applications. A key dierence is that this work considers the design of agents who directly interact with humans and attempt to improve the outcome of these interactions whereas I consider adversarial settings where an agent plans against a human adversary, but will not have any direct interaction with that human adversary in order to influence his decisions. Particularly in line with my work, Yang et al. [2011] have examined incorporating better models of human decision making within the Stackelberg framework for addressing human ad- versaries in security games. Specifically, Yang et al. [2011] have designed algorithms to compute solutions based on prospect theory [Kahneman and Tversky, 1979] and the quantal response (QR) model [McKelvey and Palfrey, 1995] to address human adversaries. Prospect theory describes human decision making as a process of maximizing ’prospect’ rather than maximizing expected value (i.e., humans do not decide according to a risk neutral utility function). Here, Yang et al. [2011] compute an optimal strategy for the security force considering that the adversary will re- spond according to the prospect theory model. The quantal response (QR) model [McKelvey and Palfrey, 1995] of human decision making is a well-founded solution concept in game the- ory derived from Nobel-prize-winning work in choice modeling theory [Nobelprize.org, 2012]. The QR model suggests that instead of strictly maximizing expected value, individuals respond stochastically in games: the chance of selecting non-optimal strategies increases as the cost of such an error decreases. In applying the QR model to security games, noise is only added to the response function for the adversary, so the security force computes an optimal strategy assuming the attacker responds with a noisy best-response. However, this approach critically depends on 40 the appropriate estimation of , which represents the amount of error or noise in the attacker’s response function. The key dierence between the work of Yang et al. [2011] and the work presented in this thesis is that Yang et al. [2011] focus on finding appropriate models of human decision making while my work focuses on finding robust strategies for humans in the absence of such models. In general, having a perfect model of human-decision-making processes would lead to the optimal strategy for security forces against human adversaries. However, finding such perfect models is a dicult task and often requires a large amount of human-decision-making data to create, which can be limited in security settings. Additionally, McCubbins et al. [2012] have recently shown that the currently existing models of human decision making are fundamentally flawed. In the absence of such data and strong models of human decision making, the robust approaches I propose in this thesis are a beneficial substitute and perform significantly better than currently existing approaches for addressing human adversaries. 41 Chapter 4: COBRA Algorithm As previously discussed, there exists a number of algorithms for Bayesian Stackelberg games that find optimal solutions considering an a priori probability distribution over possible follower types [Conitzer and Sandholm, 2006; Paruchuri et al., 2008; Kiekintveld et al., 2009; Jain et al., 2011b, 2010; Yin and Tambe, 2012]. Unfortunately, to guarantee optimality, these particular algorithms make strict assumptions on the underlying games, namely that the players are perfectly rational and that the followers perfectly observe the leader’s strategy. However, these assumptions rarely hold in real-world domains, particularly when dealing with humans. Of specific interest are security domains where there are limited resources to protect a critical set of targets [Jain et al., 2010; Pita et al., 2011; Shieh et al., 2012] – even though an automated program may determine an optimal leader (defender) strategy, it must take into account a human follower (adversary). Such human adversaries may not be expected-value maximizers, computing optimal deci- sions. Instead, their decisions may be governed by their bounded rationality [Simon, 1956] which causes them to deviate from their expected optimal strategy. Humans may also suer from limited observability of the defender’s strategy, given them a false impression of that strategy. In other words, when making decisions based on their own cognitive abilities, humans are biased due to their bounded rationality and inability to obtain complete sets of observations. Thus, a human 42 adversary may not respond with the game theoretic optimal choice, causing the defender to face uncertainty over the gamut of adversary’s actions. Therefore, in this work, the leader in a Stackelberg game must commit to a strategy consid- ering three dierent types of uncertainty: (i) adversary response uncertainty due to his bounded rationality where the adversary may not choose the expected-value-maximizing strategy, (ii) ad- versary response uncertainty due to his limitations in appropriately observing the leader strategy, and (iii) adversary reward uncertainty modeled as dierent reward matrices with a Bayesian a pri- ori distribution assumption (i.e., a Bayesian Stackelberg game). While game-theoretic optimal algorithms such as DOBSS, which are based on the standard strict game-theoretic assumptions, can handle the third type of uncertainty, these models can give a severely under-performing strat- egy when the adversary deviates because of the first two types of uncertainty. This degradation in leader rewards may be unacceptable in certain domains. To remedy this situation, I draw inspiration from robust optimization methodology, in which the decision maker optimizes against the worst outcome over some uncertainty [Aghassi and Bertsimas, 2006; Nilim and Ghaoui, 2005]. I also draw inspiration from psychological support theory for human decision making when humans are given a discrete set of actions and an un- known probability function over those actions [See et al., 2006b; Tversky and Koehler, 1994]. In the presented Stackelberg problem, the leader will make a robust decision by considering that the follower, who may not follow expected-value-maximizing rationality, could choose a strategy from his range of possible responses that degrades the leader expected value the most or that he could choose a strategy that is based on his limited observations. To that end, I introduce a mixed-integer linear program (MILP), COBRA (Combined Ob- servability and Rationality Assumptions), that builds on the Bayesian Stackelberg game model 43 in DOBSS. This MILP continues to handle adversary reward uncertainty in the same fashion as DOBSS. Along with handling reward uncertainty, it also addresses the uncertainty that may arise from human imprecision in choosing the expected optimal strategy due to bounded rationality and limited observations. Namely, it introduces the idea of robust responses to-optimal follower re- sponses into DOBSS and Stackelberg games in general. It also utilizes the concept of anchoring biases to protect against limited observation conditions, handling observational uncertainty. I first describe in depth the key ideas behind COBRA and then incrementally define the MILP that uses them. 4.1 Key Ideas The two main ideas in COBRA are addressing boundedly rational opponents and anchoring biases those opponents may have. 4.1.1 Bounded Rationality COBRA assumes that the follower is boundedly rational and may not strictly maximize expected value. As a result, the follower may select an-optimal response strategy, i.e., the follower may choose any of the responses within-expected-value of his optimal strategy. This choice may be caused by a variety of reasons, but COBRA attempts to guard against the choices that fall within this-bound of the optimal response. More specifically, given multiple possible-optimal responses, the robust approach is to as- sume that the follower could choose the one that provides the leader the worst expected value – not necessarily because the follower attends to the leader reward, but to robustly guard against 44 the worst-case outcome. In an adversarial setting, handling the worst-case outcome may be in the best interest of the leader. This worst-case assumption contrasts those of other Stackelberg solvers which assume the follower will play a strong Stackelberg equilibrium (choosing a strategy that favors the leader in the case of a tie) [Conitzer and Sandholm, 2006; Paruchuri et al., 2008], making COBRA novel to address human followers. 4.1.2 Anchoring Theory Anchoring is a cognitive bias where humans rely too heavily (i.e., anchor) on a trait or piece of information when making decisions. Specifically, once a human has anchored on some piece of information, their decision making is biased toward adjusting or interpreting other information to reflect the anchored information. Tversky and Koehler [1994] developed a theory of subjective probability known as support theory, which examines one particular type of anchoring bias where subjects anchor on the way information is presented to them. Research based on support theory has shown that when, given no information about the occurrence of a discrete set of events, hu- mans will tend to assign an equal weight to the occurrence of each event (a uniform distribution) [Fox and Clemen, 2005; Fox and Rottenstreich, 2003; See et al., 2006b]. This is also referred to as giving full support to the ignorance prior where the ignorance prior represents a subjects baseline belief (i.e., in this case the uniform distribution over events) [Fox and Rottenstreich, 2003]. It has been shown through extensive experimentation that humans are particularly susceptible to giving full support to this ignorance prior (uniform distribution) before they are given any information and that, once given information, they are slow to update away from this assumption [Fox and Clemen, 2005; Fox and Rottenstreich, 2003; See et al., 2006b]. Thus they leave some 45 support, 2 [0::: 1], on the ignorance prior and the rest, 1, on the occurrence they have actually viewed. As humans become more confident in what they are viewing, this bias begins to diminish, decreasing the value of. Models have been proposed to address this bias and predict what probability a human will assign to a particular event e from a set of events E based on the evaluative assessment (i.e., as- sessment based on events actually viewed) they have made for the occurrence of that event. Let e represent a particular event, Ene represent the remaining events possible, let P(e); P(Ene) be the real probabilities and P(e 0 ); P((Ene) 0 ) represent the probability a human assigns to event e and Ene respectively. One model [Fox and Rottenstreich, 2003] defines the human estimated proba- bilities with the following ratios: P(e 0 )=P((Ene) 0 ) = (jej=jEnej) (P(e)=P(Ene)) 1 . Here,jej=jEnej is the ratio in the case of uniform probabilities and represents the ignorance prior. The value indicates the relative contribution of these two sources of information. Note that as approaches 1, the estimated probability converges on the uniform assumption, while when approaches 0, it is closer to the true probability distribution. Research suggests that as people gain more rele- vant knowledge they will give less support to the ignorance prior and more support to evaluative assessment thus decreasing the value of [Fox and Clemen, 2005; Fox and Rottenstreich, 2003; See et al., 2006b]. An alternative model assumes estimated probabilities are directly calculated using a simple linear model [Fox and Clemen, 2005; Fox and Rottenstreich, 2003; Tversky and Koehler, 1994]: P(e 0 ) = (1=jEj) + (1) P(e). I commandeer this anchoring bias for Stackelberg games to determine how a human follower may perceive the leader strategy. For example, in the game shown in Table 2.1, suppose the leader strategy was to play a with a probability of 0:8 and b with 0:2. Anchoring bias would predict that in the absence of any information ( = 1), humans will 46 assign a probability of 0:5 to each of a and b, and will only update this belief (alter the value of ) after observing the leader strategy for some time. Although these may not be the only possible models for determining anchoring bias, they are standard in the related literature [Fox and Rottenstreich, 2003; Tversky and Koehler, 1994] and the linear model is ideal since the odds form model is not easily representable in an MILP. As an alternative approach I could use Bayesian updating to predict how humans will perceive the probability of each event from a set of events after obtaining some observations. However, there is more support in the literature that humans act according to subjective probability and anchoring biases rather than performing Bayesian updating when evaluating evidence [Fox and Clemen, 2005; Fox and Rottenstreich, 2003; Kahneman and Tversky, 1972; See et al., 2006b; Tversky and Koehler, 1994]. Also, using Bayesian updates requires tracking which observations the humans have received, while in the real world defenders will not be aware of which days adversaries take observations and which days they do not. Anchoring bias is more general in that it allows the defender to work with just an estimate of how many observations she believes an adversary will take rather than which specific observations he takes. Specifically, a value is assigned to based on how much evidence the defender thinks the human will receive. If the human adversary is expected to observe the defender’s policy frequently and carefully then will be low while if the defender suspects that the adversary will not have many observations of her policy, will be high. 47 4.2 Robust Algorithm COBRA(;) is the new algorithm I introduce in this thesis. To introduce it in steps, I will first introduce two simplified versions, which I will refer to as COBRA(0,) and COBRA(, 0). CO- BRA(0,) deals only with bounded rationality and COBRA(, 0) deals only with observational uncertainty. After introducing each of these pieces individually I will combine them into a single algorithm that can handle both types of uncertainty which I refer to as COBRA(;). For each of these algorithms and represent two parameters that can be adjusted. 4.2.1 COBRA(0,) COBRA(0, ) considers the case of a boundedly-rational follower, where COBRA(0, ) maxi- mizes the minimum expected value it obtains from any-optimal response from the follower. In the following MILP, I use the same variable notation as in DOBSS from Chapter 2 Section 2.4: 48 max X 2 p V s:t: X i2S x i = 1 (4.1) x i 2 [0::: 1] 8 i2 S (4.2) X j2S q j = 1 8 2 (4.3) X j2S h j 1 8 2 (4.4) q j h j 8 2 ; j2 S (4.5) q j ; h j 2f0; 1g 8 2 ; j2 S (4.6) 0 a X i2S ( i ; j ; ) x i (1 q j ) M 8 2 ; j2 S (4.7) (1 h j ) a X i2S ( i ; j ; ) x i + (1 h j ) M 8 2 ; j2 S (4.8) V M (1 h j ) + X i2S ( i ; j ; ) x i 8 2 ; j2 S (4.9) a 2R 8 2 (4.10) As before, the variables q j identify the optimal strategy for follower type with a value of a in Constraints 4.3 and 4.7. Variables h j represent all-optimal strategies for follower type ; Constraint 4.4 allows selection of more than one-optimal strategy per follower type. Constraint 4.8 ensures that h j = 1 for every action j such that a P i2S ( i ; j ; ) x i <, since in this case the middle term in the inequality is less than and the left inequality is then only satisfied if h j = 1. This robust approach required the design of a new objective and an additional constraint. 49 Constraint 4.9 helps define the objective value against follower type , V , which must be lower than any leader expected value for all actions h j = 1, as opposed to the DOBSS formulation which has only one action q j = 1. Setting V to the minimum leader expected value allows COBRA(0;) to robustly guard against the worst-case scenario. 4.2.2 COBRA(; 0) COBRA(; 0) considers the case where the human follower is perfectly rational, but faces limited observations. COBRA(; 0) draws upon the theory of anchoring biases mentioned previously to help address the human uncertainty that arises from such limited observation. It deals with two strategies: (i) the real leader strategy (x) and (ii) the perceived strategy by the follower (x 0 ), where x 0 is defined by the linear model presented earlier. Thus, x i is replaced in Constraint 4.15 with x 0 i and x 0 i is accordingly defined as x 0 i = (1nj j) + (1) x i . The justification for this replacement is as follows. First, this particular constraint ensures that the follower maximizes his expected value. Since the follower believes x 0 to be the leader strategy then he will choose his strategy according to x 0 and not x. Second, given this knowledge, the leader can find the follower’s responses based on x 0 and optimize her actual strategy x against this strategy. Since x 0 is a combination of the support for x and the support toward the ignorance prior, COBRA(; 0) is able to find a strategy x that will maximize the leader’s expected value based on the relative contribution of these two sources of support. For consistency among these new approaches I use the same objective introduced in COBRA(0;). The new MILP then is as follows: 50 max X 2 p V s:t: X i2S x i = 1 (4.11) x i 2 [0::: 1] 8 i2 S (4.12) X j2S q j = 1 8 2 (4.13) q j 2f0; 1g 8 2 ; j2 S (4.14) 0 a X i2S ( i ; j ; ) x 0 i (1 q j ) M 8 2 ; j2 S (4.15) V M (1 q j ) + X i2S ( i ; j ; ) x i 8 2 ; j2 S (4.16) a 2R 8 2 (4.17) x 0 i = (1nj j) + (1) x i 8 i2 S (4.18) 4.2.3 COBRA(;) COBRA(;) is an MILP that combines both a bounded rationality assumption and an ob- servational uncertainty assumption. This is achieved by incorporating the alterations made in COBRA(; 0) and COBRA(0;) into a single MILP. Namely, COBRA(;) includes both the parameter and the parameter from COBRA(0;) and COBRA(; 0) respectively. The MILP that follows is identical to COBRA(0;) except that in Constraints 4.24 and 4.25, x i is replaced with x 0 i as it is in COBRA(; 0). The justification for this replacement is the same as in COBRA(; 0). The new MILP then is as follows: 51 max X 2 p V s:t: X i2S x i = 1 (4.19) x i 2 [0::: 1] 8 i2 S (4.20) X j2S q j = 1 8 2 (4.21) X j2S h j 1 8 2 (4.22) q j h j 8 2 ; j2 S (4.23) q j ; h j 2f0; 1g 8 2 ; j2 S (4.24) 0 a X i2S ( i ; j ; ) x 0 i (1 q j ) M 8 2 ; j2 S (4.25) (1 h j ) a X i2S ( i ; j ; ) x 0 i + (1 h j ) M 8 2 ; j2 S (4.26) V M (1 h j ) + X i2S ( i ; j ; ) x i 8 2 ; j2 S (4.27) a 2R 8 2 (4.28) x 0 i = (1nj j) + (1) x i 8 i2 S (4.29) 4.2.4 Complexity It has been shown that finding an optimal solution in a Bayesian Stackelberg game is NP-hard [Conitzer and Sandholm, 2006] and thus DOBSS, COBRA(; 0), COBRA(0;), and COBRA(;) are MILPs that face an NP-hard problem. A number of eective solution packages for MILPs can be used, but their performance depends on the number of integer variables. DOBSS and 52 COBRA(; 0) considerj jjj integer variables, while COBRA(0;) and COBRA(;) double that. MAXIMIN on the other hand is a linear programming problem that can be solved in poly- nomial time. Thus it is anticipated that MAXIMIN will have the lowest running time per problem instance, followed by DOBSS and COBRA(; 0) with COBRA(0;) and COBRA(;) close be- hind. However, I will show in Section 4.4.5, this was not observed in practice. 4.3 Equivalences Between Models In this section I suggest and prove equivalences between COBRA(;) and DOBSS under certain conditions. First I will demonstrate how DOBSS can be reformulated with the same objective as COBRA(; 0). After this reformulation I will show how altering the parameters and can cause COBRA(;), and DOBSS to produce identical results (i.e., identical mixed strategies). Observation 1. When = 0 and = 0 then COBRA(;) and DOBSS are equivalent Proof. It follows from the definition of x 0 i that when = 0 then x 0 i = x i since the follower is assumed to once again perfectly observe and believe the leader strategy x i . Note that if = 0 the inequality in Constraint 4.25 of COBRA(;) is the same expression as the inequality in Constraint 4.24 with h j substituted for q j . Since the objective of COBRA(;) is to maximize the leader expected value this means that q j will be selected as the follower’s optimal response that maximizes the leader expected value (i.e., a strong Stackelberg equilibrium) and h j will be set to the same. Introducing additional h k = 1 where k , j would only serve to reduce the expected value and thus would not be an optimal solution to the MILP. I will show that DOBSS and COBRA(0, 0) attain the same optimal objective function value. 53 To show that solution to COBRA(0,0) solution to DOBSS, consider (q; z; a) a feasible solu- tion for DOBSS. We define ¯ x i = P j2S z i j ; ¯ q = ¯ h = q; ¯ a = a, and ¯ V = P i2S P j2S ( i ; j ; ) z i j . From Constraints 2.7, 2.8, 2.10, and 2.14 in DOBSS we can show that z i j = 0 for all j such that q j = 0 and thus that ¯ x i = z i j for all j such that q j = 1. This implies that ¯ V = P i2S ( i ; j ; ) ¯ x i for the j such that q j = 1 in DOBSS and it is then easy to verify that ( ¯ x; ¯ q; ¯ h; ¯ a; ¯ V) is feasible for COBRA(0, 0) with the same objective function value of (q; z; a) in DOBSS. For solution to DOBSS solution to COBRA(0,0), consider (x; q; h; a; V) feasible for CO- BRA(0,0). Define ¯ q = q; ¯ z i j = x i q j , and ¯ a = a. Then we can show that (¯ q; ¯ z; ¯ a) is feasible for DOBSS by construction. For COBRA(0,0), since q j h j in Constraint 4.22 and as explained in the optimal solution h j will equal q j it follows that V P i2S ( i ; j ; ) x i for the j such that q j = 1. This implies that V P i2S P j2S ( i ; j ; ) ¯ z i j and that the objective func- tion value of (¯ q; ¯ z; ¯ a) in DOBSS is greater than or equal to the objective value of (x; q; h; a; V) in COBRA(0,0). The key implication of the above observation is that when = 0, COBRA(;) loses its robustness feature, so that once again when the follower faces a tie, it selects a strategy favoring the leader, as in DOBSS. Based on this observation, the remaining observations presented in this thesis about COBRA(;) can be generalized to DOBSS accordingly. Observation 2. When is held constant, the optimal expected value COBRA(;) can obtain is decreasing in. Proof. Since Constraint 4.25 in COBRA(;) makes h j = 1 when that action has a follower expected value between (a ; a ], increasing would increase the number of follower strategies 54 set to 1. Having more active follower actions in Constraint 4.26 can only decrease the minimum value of V . Observation 3. Regardless of, if 1 2 > Pj ( i ; j ; )j 8 i; j; , where P is the greatest absolute opponent payo, then COBRA(;) is equivalent to MAXIMIN. Proof. Note thatja j in COBRA(;) P. The proof simply needs to show that the leftmost inequality of Constraint 4.25 in COBRA(;) implies that all h l j must equal 1. This would make COBRA(;) equivalent to MAXIMIN. We know from the observation itself thatP P i2S ( i ; j ; ) x i . Suppose some h j = 0, then we can reorganize the leftmost inequality of Constraint 4.25 to state that P i2S ( i ; j ; ) x i a . From the observation itself we can then obtain a < P 2P =P and thus we haveP P i2S ( i ; j ; ) x i <P a contradiction. Although DOBSS and COBRA(0;) make dierent assumptions about the follower’s re- sponses, it can be shown that in a zero-sum game their optimal solutions become equivalent. This of course is not true of general sum games. In a zero-sum game the rewards of the leader and follower are related by ( i ; j ; ) = ( i ; j ; ) 8 i2 S ; j2 S ; 2 : It is well known that maximin strategies constitute the only natural solution concept for two player zero-sum games [von Neumann, 1927]. As DOBSS is an optimal Stackelberg solver it follows that it is equivalent to a maximin strategy for zero-sum games. Observation 4. COBRA(0;) is equivalent to a maximin strategy for zero-sum games. 55 Proof. Given the maximin solution to a zero-sum game we define the expected value for any strategy of the follower, j 2 , to be W j for the leader and W j for the follower. Assuming that the follower’s strategies are ordered in descending order in terms of expected value for the follower and there are m such strategies, the maximin solution yields W 1 W 2 ::: W m for the follower and W 1 W 2 ::: W m for the leader. COBRA(0, ) is an algorithm that assumes the attacker will choose a strategy within of his maximum expected value (in this case W 1 ) and of these-optimal strategies it attempts to maximize the expected value for the worst- case outcome. As seen by W 1 ::: W m the minimum expected value possible for the leader is the follower’s optimal strategy and any-deviation will only result in a higher expected value for the leader. It follows that the maximin strategy already maximizes the expected value for the worst-case outcome of any-optimal response by the follower. Thus, COBRA(0;) yields a strategy that is equivalent to maximin in zero-sum games. 4.4 Experiment Purpose, Design, and Results 4.4.1 Purpose of this Study I sought to investigate the performance of several previously existing approaches against my new robust approach (COBRA(;)) under several variables. In particular, I examine the performance of COBRA(;) against DOBSS and the two baseline approaches (MAXIMIN and UNIFORM) described in Section 2.4. Here, performance is measured by the average expected value obtained by a security force against the decisions of human adversaries. These experiments were setup to examine three crucial variables of real-world domains: i) reward structure, ii) observation 56 condition, and iii) parameter settings for and. In the following sections I will describe each of these variables in depth. The goal of COBRA(;) was to increase the performance of a security force against human adversaries by addressing the bounded rationality that humans may exhibit and the limited obser- vations they may experience in many settings. To that end, experiments were set up where human subjects would play as followers (adversaries) against each strategy with varying observability conditions under dierent reward structures. It is not possible to prove optimality against human adversaries who may deviate from the expected optimal responses and thus I rely on empirical validation through experimentation. In addition to examining the performance of previously ex- isting approaches with COBRA(;), I also examine the runtime performance of COBRA(;) against each of the previous approaches. 4.4.2 Experimental Design In order to examine these three variables (reward structure, observation condition, algorithm pa- rameters) I constructed a game inspired by the security domain at LAX [Pita et al., 2008] de- scribed in Section 2.7, but converted it into a pirate-and-treasure theme. The domain has three pirates – jointly acting as the leader – guarding 8 doors, and an individual human subject would act as a single adversary. The 8 doors model the 8 terminals found at LAX. The interface can be seen in Figure 4.1. Although COBRA(;) allows for multiple follower types (a Bayesian Stackelberg game) each subject was modeled as a single follower type defined by the reward structure they were given. The subject’s goal was to steal a treasure from behind a door without getting caught. Each of the 8 doors would have a dierent reward and penalty associated with it for both the subjects 57 as well as the pirates. For instance, as shown in Figure 4.1, door 4 has a reward of 3 and a penalty of -3 for the subject and a reward of 1 and a penalty of -5 for the pirate. If a subject chose a door that a pirate was guarding, the subject would incur the subject penalty for that door and the pirate would receive the pirate reward for that door, else vice-versa. Going back to the previous example, if the subject chose door 4 and a pirate was guarding that door then the subject would receive -3 and the pirate would receive 1. This setup led to a Stackelberg game with 8 3 = 56 leader actions, and 8 follower actions. 4.4.2.1 Participants There are two dierent experimental evaluations that were conducted in this study. In the first experimental evaluation there were 178 participants consisting of Engineering undergraduate and graduate students at the University of Southern California. Of the subjects in the first experi- mental evaluation, 81% were men and ages ranged from 17 to 32 (M = 22, SD = 3). In the second experimental evaluation there were 40 participants consisting of Engineering undergrad- uate and graduate students at the University of Southern California. Of the subjects in the second experimental evaluation, 80% were men and ages ranged from 18 to 28 (M = 22, SD = 2). 4.4.2.2 Reward Structure One of the variables examined is the reward structure of the pirate-and-treasure domain. De- pending on the reward structure human choices could vary vastly. For instance, in some reward structures there may be only a single action that obtains the highest expected value possible, while in a dierent reward structure there may be multiple dierent actions that obtain the high- est expected value possible. For this study I examine reward structures similar to those used in 58 the security domain at Los Angeles International Airport (LAX) described previously [Pita et al., 2008]. I constructed four dierent reward structures for the 8-door 3-pirate domain described. These reward structures can be found in the Appendix under Section B, Tables B.1-B.4. There are three key features in these reward structures. First, the reward scale is similar to that used at LAX to determine the payos for both the leader and the follower. Namely, rewards range from 1 to 10 and penalties range from -10 to -1. Second, these reward structures meet the model criteria of what are known as security games as described in Chapter 2 Section 2.6, which are used in domains such as LAX and FAMS Jain et al. [2010]. Finally, in addition to the reward scale and to ensuring that the requirements of a security game were met, I also wanted to examine reward structures where the follower’s small-deviations from the strong Stackelberg equilibrium (SSE) were significantly harmful to the leader (such arbitrarily small -deviations model cases where the follower does not break ties in favor of the leader). These reward structures are particularly interesting for the new robust method of COBRA(;) since it specifically guards against such potentially harmful deviations. Reward structure four is the baseline case, a zero-sum reward structure, where deviations from the follower’s optimal strat- egy based on a SSE assumption are only better for the leader. In the other three reward structures, the defender’s average expected value for an-deviation – for arbitrarily small – from a SSE is held between -1.30 and -1.96; however, the worst-case expected value for a deviation becomes progressively worse for each reward structure. In reward structure three the worst-case deviation is -3.16, in reward structure two it is -4.21, and finally in reward structure one it is -4.56. Thus, for each of these three reward structures, follower’s deviations from optimal play based on a SSE assumption can lead to potentially large degradations in the defender’s expected value. 59 As stated before, Figure 4.1 shows the interface used to convey the reward structures to the subjects. The worst-case outcome for deviation from SSE is progressively better from reward structure one to reward structure three. The key question is whether there is any systematicity in human subject’s deviations from SSE under dierent conditions and if COBRA(;) can capture them suciently to mitigate their impact. If so, then it is expected that the robust model will provide the largest benefit in reward structure one and the least benefit in reward structure three. In reward structure four it has been shown that, given no observational uncertainty, COBRA(;) is equivalent to the optimal maximin strategy and thus there will be no benefit due to this robust approach. However, in all four reward structures it is expected that COBRA(;)’s method for handling observational uncertainty will provide benefits when observational capabilities are low. 4.4.2.3 Observability Conditions A second variable in this experimental design is dierent observability conditions for human ad- versaries. Security ocials, such as those at LAX, are interested in the observational capabilities of their adversaries and how this will aect their decisions. In the real-world, some adversaries may be able to take many observations before deciding to act while others may end up having to act with very little information. It is important to understand how this can aect the decisions and actions humans may take. There are four separate observability conditions that were examined. First I explain what an observation is exactly. Assume that time can be discretized into rounds and on each round the pirates will choose three doors to guard according to their current mixed strategy. For example, examining the UNIFORM strategy, on each round the guards will choose three doors uniformly at random to guard for that round. A single observation then, consists of seeing where the guards 60 have stationed themselves for a single round and then moving on to the next round where the guards will once again reposition themselves according to their mixed strategy. An example of an observation can be seen in Figure 4.2. On a separate window the observations the subjects had received in the current game were printed for them to refer back to. For instance, after five observations the separate window may list something like [1,2,4][2,4,5][2,6,7][3,7,8][1,3,6] where each triplet is the three doors the pirates were stationed at in each of the five rounds. On each round the subject will either receive an observation or will be asked to choose a door to attack. The four dierent observation conditions tested were: (i) The subject does not get any obser- vations (a 1 round game), (ii) the subject gets 5 observations (a 6 round game), (iii) the subject gets 20 observations (a 21 round game), and (iv) the subject gets unlimited observations – sim- ulated by revealing the exact mixed strategy of the pirates to the subject. This final condition is also a 1 round game, but the mixed strategy the pirates used to select their 3 doors is displayed and the subjects are then allowed to choose which door they will attack on this first round. The mixed strategy is revealed as the marginal distribution of guards over the 8 doors. Specifically, for each door the subject is given the probability that any guard will appear on that door for any given round. As shown in Figures 4.1 and 4.2, subjects were given full knowledge of their rewards and penalties and those of the pirates in all situations. In each game the subject observed the pirates’ strategy under the current observability condition, reward structure, and strategy and then was allowed to make his decision on the final round of the game. After making his decision the subject was informed whether he was successful or whether he was caught. 61 Figure 4.1: Game Interface Figure 4.2: Single Observation 4.4.2.4 Algorithms and Parameters Finally, the last variable examined was the performance of COBRA(;) against human sub- jects given adjustments to its and parameters under dierent settings for the first two vari- ables in an attempt to understand how to better deal with human adversaries. These experiments 62 compare DOBSS, COBRA(0;), COBRA(;), MAXIMIN, and UNIFORM. I chose to include COBRA(0;) in these experiments to demonstrate the value of having both observational uncer- tainty and bounded rationality assumptions in an algorithm over having just one. It is important to include either COBRA(; 0) or COBRA(0;) to examine whether observational uncertainty, bounded rationality, or the combination of both results has the largest aect on the results. Based on the results in all four conditions for COBRA(0;) the need to address both bounded ratio- nality and observational uncertainty was found to be important. Furthermore, the second set of experiments were conducted to explore the impact of. The value of in an application should reflect the assumptions regarding the uncertainty in adversary reward and the precision with which the human adversary decides between the dierent actions. While determining the human adversary’s precision is dicult in general, in these exper- iments the value of was determined based on three factors: (i) > 0 based on the assumption that subjects playing these games were not precisely computing expected utilities, (ii) had to be set to a value to produce qualitatively dierent leader strategies than created by DOBSS and MAXIMIN to help gain a clear understanding of how aects the results against human subjects, and (iii) had to be held constant across the four games. Hence I chose = 2:5, which lead to 3 to 4 actions in the subject’s-optimal set. This is about the halfway point given these exper- iments had 8 total choices of actions. DOBSS assumes a single action choice to the adversary and MAXIMIN makes a worst-case assumption. In some settings a higher or lower selection of may be appropriate. Finding a more precise method for selecting is left for future studies. Unlike deciding ", a single choice for will not hold across all reward structures. First, could be expected to vary with observability conditions. Second, even for a fixed observability 63 condition, identical values of across reward structures is not appropriate. Included in the Ap- pendix under Section F are tables for how COBRA(,2.5) changes as changes from 0 to 1. Notice that in some reward structures changing may not necessarily change the mixed strategy for the leader as in Table F.4. For this reason it is necessary to examine each reward structure individually. Within each reward structure, I tried two dierent techniques to choose, one with a fixed and one with a variable, in order to compare the impact of variable more clearly. Since the value of is used to balance the amount of observed information and a priori bias information that the adversary incorporates in his assumption of the leader strategy, this parameter should be related to the amount of observations made by the follower. Thus, the technique with variable is obviously the more standard version. Clearly when the follower has unlimited information = 0 (the follower correctly estimates the leader strategy) and when he has no observations = 1 (the follower uses the a priori bias). Less straight forward is how to set an value when the follower has 5 or 20 observations. I follow two methods of adjusting: The first is to conduct r trial experiments with human adversaries using COBRA(0,") with 0, unlimited, and n observations. In these experiments n is either 5 or 20. Given the choices made by each subject in the trial experiments, r subject expected values are col- lected for each observation condition. Let corr n;0 be the correlation between the r subject expected values for the n observation condition and the 0 observation condition. Similarly let corr n;u be the same with the unlimited observation condition. is set for n observations as = corr n;0 =(corr n;0 + corr n;u ). If the results were more correlated with the unobserved condition this would make higher and otherwise it would make lower. It is assumed 64 that having more observations would lead to results that were more correlated with the un- limited observation condition and having less observations would lead to results that were more correlated with the unobserved condition. I use this form of adjusting in reward structures 1 and 2. The second method to set uses fixed arbitrary values: = :75 for the 5 observation condition and =:25 for the 20 observation condition. This simple method was created to evaluate the necessity of conducting trial runs to determine the values experimentally. I use this second method in reward structures 3 and 4. Finding an exact method to select in these conditions remains an issue for future studies. Before running this study there was no evidence to suggest either method attempted would work best. However, I hoped to gain some insight into choosing based on the results obtained from both approaches. Section 4.4.4 provides more analysis for choosing. The second technique to selecting is to assume a constant , leading to a version of COBRA(,") that I will refer to as COBRA(C,") (COBRA(,") with constant). For COBRA(C,"), was set to the same value as the 5 observation condition from the four reward structures with the expectation that it would perform poorly in higher observation conditions since it was not appropriately adjusted. The values chosen for COBRA(,") in each of the reward structures are summarized in Table 4.1. In Section 4.4.4 I will present experimental results for many addi- tional settings of the parameter in both the unobserved and unlimited observation conditions. Based on these settings I will introduce a third, better performing method for determining that generalizes to all four reward structures and is more practical in the field. 65 Unobserved 5 20 Unlimited Structure One 1 .37 .03 0 Structure Two 1 .54 .41 0 Structure Three 1 .75 .25 0 Structure Four 1 .75 .25 0 Table 4.1: values used in each observation condition for a given reward structure 4.4.2.5 Experimental Procedure For the first experiment, each of the 96 game settings (four reward structures, six algorithms, and four observability conditions) were played by 40 unique subjects (i.e., in total there were 2480 total trials). There are fewer than 3840 trials however because some are duplicates. For example, the unobserved condition for a particular reward structure need only be played by 40 subjects as opposed to 240 subjects (one set of 40 for each algorithm). The choices made by those 40 subjects could then be used to compare all six algorithms since the subjects were not given any information about a particular algorithm in advance for this condition. Thus the choices made were irrelevant of the algorithm used. Also notice that in some observation conditions a few of the algorithms yielded the same strategy (e.g., COBRA(,") and COBRA(C,") in the five observation condition used the same" and parameters). The 96 potential game settings were broken into two separate groups of 48 game settings. The first group of 48 game settings consisted of all combinations of game settings for reward structures 1 and 2 (i.e., 2 reward structures, six algorithms, and four observability conditions). Similarly, the second group consisted of the same for reward structures 3 and 4. All experiments were first conducted for group 1 and then group 2. The following procedure was used for both groups. Given these 48 game settings, each subject played a total of 14 unique games. For now I will omit the 2 game settings where the subjects play the unobserved observation condition for each 66 reward structure. This leaves 46 potential game settings. For every 10 subjects, 14 games were chosen completely at random without replacement from the 46 possible remaining game settings. This single random ordering was played by all 10 subjects. Once a particular game setting was played by 40 subjects it was removed completely from the set of possible game settings. If less than 14 potential game settings remained, the remaining games were chosen at random from the completed game settings, but not recorded. For example, if there were only 12 potential game settings that did not have 40 total subjects, then 2 game settings would be chosen from the other 44 game settings that did have 40 total subjects, but the results for these 2 games would not be recorded. The first 40 subjects played the 2 game settings that I previously omitted. In this case, the first two games played were always these 2 game settings (i.e., the unobserved observation condition for each reward structure). This was done so they would not have viewed any strategies that could influence their anchoring bias. After these 2 games, the next 12 were chosen for every 10 subjects using the same procedure outlined above. I presented the games in random orderings in an attempt to reduce ordering aects. Subjects were not allowed to participate in this study more than once. The actual experiments were conducted in a campus oce on two standard desktops. When subjects came in they were presented with the instructions seen in the Appendix under Section G.1. Subjects were given an unlimited amount of time to study the instructions and ask questions before beginning to ensure they understood how the game was played. Once subjects began they were also given an unlimited amount of time to complete the game. 67 For each game, the objective of a subject was to earn as many points as possible. The subject was allowed to choose a single door, based on the current reward structure and observation con- dition, that they believed was unguarded and once a door was chosen that game was over and the subject played the next game. As stated previously, at the end of each individual game the subject was informed whether he was successful or not. Starting with a base of US $8.00, each reward point within the game was worth US $0.15 for the subject and each penalty point deducted US $0.15. This was incorporated to give the subjects incentive to play as optimally as possible. For a given algorithm the leader expected value was computed for each follower action (i.e., for each choice of door by subject). Then, the average expected value was calculated for a given algorithm using the actual door selections from the 40 subject trials. For the second experiment, there were only 12 game settings in total. While I will describe in detail these 12 game settings in Section 4.4.4, for these experiments I used the exact same setup described above, but the subjects only played 12 games instead of 14. Again, each set of 10 subjects was presented one random ordering of the 12 games. 4.4.3 Experimental Results Figure 4.3(a) shows the average leader expected value for the first reward structure, with each data-point averaged over 40 human responses. Figures 4.3(b), 4.3(c), and 4.3(d) show the same 68 for the second, third, and fourth reward structures 1 . In both figures, the x-axis shows the obser- vation condition for each strategy and the y-axis shows the average expected value each strat- egy obtained. For example, examining Figure 4.3(a) in the unlimited observation condition, COBRA(C,") scores an average leader expected value of -0.33, whereas DOBSS suers a 663% degradation of expected value, obtaining an average score of -2.19. (a) Reward structure one (b) Reward structure two (c) Reward structure three (d) Reward structure four Figure 4.3: Average leader expected value 1 The reason all strategies obtain a negative average is due to the lack of enough resources (or guards) in this setting, but that is where randomization strategies have the most impact. The reason COBRA(,") was able to obtain a positive average expected vale in the unobserved condition is that it placed all its resources on a small subset of doors (a more deterministic strategy) assuming that humans would choose these doors based on their belief of the ignorance prior (uniform distribution). In practice the humans did indeed play according to this expectation and thus the expected values obtained were much higher. Examining Tables D.1-D.4 in the Appendix under Section D, it is clear that while most expected values are negative, in the unobserved condition for COBRA(,") (seen as COBRA(1,2.5)) there are a few distinct doors that obtain very high expected values, which correspond to the doors chosen by most human subjects in these experiments for this condition. 69 4.4.3.1 Key Observations I provide my key observations first, then provide statistical significance tests in the following section and later provide a deeper analysis. The main observation from Figure 4.3 is that the COBRA(,") algorithm has a performance that is superior to the theoretically optimal DOBSS algorithm and baseline approaches against human subjects. I can breakdown this main observa- tion into observations of the following trends: 1. Considering observational uncertainty is important: When creating algorithms for leader strategies in Stackelberg games it is important to address biases that may arise from ob- servational uncertainty. Focusing first on the unobserved and 5 observation conditions, COBRA(,") and COBRA(C,") obtain much better results demonstrating the benefit of incorporating an anchoring bias. In the 20 and unlimited observation conditions it can be seen that COBRA(,") and COBRA(C,") can still provide benefits, however, in these cases it is assumed that observational uncertainty is low so the eect of anchoring biases begin to diminish. As can be seen in Figure 4.3, COBRA(,") performs better than DOBSS in all except only two cases (Reward structure three under 20 and unlimited observations). Be- cause of its lack of adjustment of, COBRA(C,") is seen to perform better than DOBSS in all conditions except in four total cases confined to reward structures three and four. 2. Addressing bounded rationality is an important component when designing algorithms that perform against humans: Examining COBRA(0,") specifically, it can be seen in the 20 and unlimited observation conditions for reward structure one and two that addressing bounded rationality provides improvements over DOBSS. In the lower observation conditions we 70 begin to see how observational uncertainty becomes a larger factor making bounded ra- tionality a secondary issue. This follows from the first observation presented. In reward structure four, a zero-sum game, COBRA(0,") and DOBSS are exactly equivalent, so no gains are expected. In reward structure three, where deviations from SSE were least harm- ful of the remaining reward structures, there is some dierence between COBRA(0,") and DOBSS, but not as much as in reward structures one and two. In general though, these results confirm the expected result that humans do not strictly play the game theoretic opti- mal strategy. Thus, it is important to address deviations from this theoretic optimal strategy based on bounded rationality to prevent what could potentially be significant losses. 3. COBRA(C,") surprisingly outperforms COBRA(,") under high observation conditions in some reward structures: This unexpected result lead to the subsequent experiments re- ported in Section 4.4.4 and to a more ecient heuristic for selecting. 4.4.3.2 Statistical Significance The main observation in the previous section critically depends on significant dierences between COBRA(,") and the remaining strategies. I chose to employ more robust statistical methods for these tests in order to overcome limitations with the data set. These limitations include a non- normal distribution (due to a very small number of discrete choices as opposed to continuous or near continuous choices) and high variance. Having a normal distribution is an important assumption of traditional statistical tests such as the classic T-test. For the statistical significance tests I used a one-way Brunner-Puri test [Brunner et al., 1999] for repeated observations in the unobserved condition and I used Yuen’s test for comparing 71 trimmed means [Yuen, 1974] in the 5, 20, and unlimited observation conditions. In the un- observed condition all structures were treated separately, however, in the 5, 20, and unlimited observation conditions reward structures one and two were combined into a single data set. For an in depth discussion of this decision and also why these statistical tests were chosen please see the Appendix, Section A. In general, given the nature of the data — discrete rather than contin- uous distribution of values and non-normal distributions — it can be dicult to obtain statistical significance without significantly larger data sets. Yet, these results do achieve statistical signif- icance in key cases demonstrating the eectiveness of the COBRA(,") strategy, as summarized below. Conclusions regarding unobserved condition: Looking first at the unobserved condition, COBRA(,") obtained statistical significance against DOBSS in reward structures one, two, and three with a maximum p-value of .04. Since reward structure four is a zero-sum game I reiterate that the strategy space for COBRA(;) is limited. Namely, in all observation conditions the results for the DOBSS, MAXIMIN, and COBRA(0,") algorithms are identical. Thus, the only way to alter the strategy based on this robust method is through the use of. Although the results are in favor of COBRA(,") in this reward structure, in the unobserved condition it was not possible to achieve statistical significance against DOBSS/COBRA(0,")/MAXIMIN without a larger data set. Against MAXIMIN and UNIFORM, COBRA(,") obtains statistical significance in reward structures one, two, and three with a maximum p-value of .04 except in reward structure one where it obtains a p-value of .098 against MAXIMIN. Overall these results demonstrate the superiority of COBRA(,") over DOBSS and simple baseline algorithms in the unobserved condition. 72 Conclusions regarding combined data from reward structures one and two in remain- ing observation conditions: In the 5, 20, and unlimited observation cases the maximum p-value obtained for COBRA(C,") versus any other strategy was .033. Given that COBRA(C,") is shown outperforming every other strategy, including COBRA(,"), under these observation conditions, this establishes that COBRA(C,") is statistically significantly better than all other strategies in these reward structures under these observation conditions. This in turn demonstrates the su- periority of the COBRA(,") algorithm as a whole in these reward structures and observation conditions since COBRA(C,") is an instantiation of COBRA(,") with a particular choice of. Conclusions regarding reward structure three in remaining observation conditions: In reward structure three for the 5 and 20 observation conditions we do not achieve statistical signif- icance between DOBSS, COBRA(,"), and COBRA(C,") making the results obtained inconclu- sive. To a certain extent this is an implication of deviations in reward structure three not being as harmful to the leader (see Section 4.4.2.2) and thus more dicult to obtain significant dierences between the strategies. However, in the unlimited observation condition we find that DOBSS is statistically significantly better than all other strategies with a maximum p-value of .036. Al- though in the unlimited condition, given the choices made for and", DOBSS outperforms the robust strategy of COBRA(;), I will later present an alternative choice for these parameters in Section 4.4.4 that is able to outperform DOBSS. Conclusions regarding reward structure four in remaining observation conditions: In reward structure four, as presented in Section 4.3, it has been shown that MAXIMIN, DOBSS, and COBRA(0,") are equivalent. In the unlimited and 20 observation conditions, given the choice for, it is also the case that COBRA(,") and DOBSS/MAXIMIN/COBRA(0,") are equivalent. Given these equivalences, our only concern is COBRA(C,"). Since COBRA(C,") is outperformed 73 in the 20 and unlimited observation conditions it provides no benefits in these cases. However, for the 5 observation condition COBRA(C,") (or COBRA(,")) is statistically significantly better than all other strategies with a maximum p-value of .005. Given these results and the statistical significance achieved in the 5 observation condition, COBRA(,") is found to be the superior strategy, even in a zero-sum game, due to its ability to handle observational uncertainty. 4.4.3.3 Analysis of Results I discuss the key implications of the observations presented in Section 4.4.3.1 and why they were reached. I include two tables for reference in the following discussion, Tables 4.2 and 4.3. Table 4.2 shows the expected values (for a subset of the algorithms tested) the leader should obtain for each door selection by the follower in reward structure one on average. For instance, if the follower selected Door 2 when playing against DOBSS the leader would expect to obtain an expected value of -.97. Obviously depending on whether there was a guard stationed there or not in any particular instance the leader would get the respective reward or penalty associated with that door, but over time the average would converge to the expectation of -.97. I have placed in bold font the predicted expected value for each of the algorithms. This predicted expected value is the expected value an algorithm expects to receive based on the as- sumptions it has made. For example, DOBSS is an algorithm that expects the follower to play a strong Stackelberg Equilibrium (SSE) strategy. This means that the follower will choose, between his highest expected value choices, the door that is also best for the leader. This SSE strategy is the expected value that has been placed in bold font for DOBSS. MAXIMIN on the other hand is an algorithm that makes a worst-case assumption and thus all doors that give the minimum expected value are placed in bold font. 74 Table 4.3 shows the percentage of times the follower chose a response that gives the leader (pirate) an expected value equivalent to or higher than the predicted expected value for the cur- rent algorithm under dierent observation conditions in reward structure one. I will refer to these responses as expected strategy(s). To clarify what I mean by expected strategy(s) I will look at COBRA(0,") as an example. As seen in Table 4.2, this algorithm expects to receive an expected value of -.36. Thus, in Table 4.3 under the unobserved observation condition it can be seen that the follower chose a door that gave the leader an expected value of -.36 or higher (an expected strategy) 65% of the time. I point out that MAXIMIN is a strategy that expects a worst-case outcome and thus all doors are expected strategies. DOBSS is on the opposite extreme, since it assumes perfectly rational play, and thus generally results in a single door being the expected strategy. COBRA(0,") and COBRA(,") fall somewhere in between these lines, where multi- ple doors within the"-optimal strategies are expected strategies, but less doors than MAXIMIN where all doors are expected strategies. Of course as shown in Section 4.3 this depends on the setting of" since a setting that is too high will yield the same result as MAXIMIN. Table 4.2 shows an important trade-o. MAXIMIN achieved a 100% match with expected strategies (Table 4.3), but it does so by making all leader expected values low (-1.63). DOBSS achieves low match with expected strategies, but its leader predicted expected value is higher (.39). COBRA(,") is in the middle of these extremes. Included in the Appendix under Sections C, D, and E respectively are: i) tables presenting the actual mixed strategies for each of the reward structures, ii) tables of the expected values for each reward structure given the strategies presented in i), and iii) tables for the percentage of times the follower chose an expected strategy in each reward structure. In each of the expected value tables I have placed in bold font the predicted expected value for each of the algorithms. I reiterate that an expected strategy is any 75 door selection that gives an expected value for the leader at least as high as the values in bold font. DOBSS COBRA(0,") MAXIMIN COBRA(.37,") COBRA(.03,2.5) COBRA-5 Door 1 -5 -4.58 -1.63 -5 -4.61 Door 2 -.97 -.42 -1.63 -.30 -.37 Door 3 .36 -.36 -1 -.30 -.37 Door 4 -1.38 -.79 -1.63 -.30 -.73 Door 5 .06 -.36 -1.63 -.30 -.37 Door 6 -1 -.86 -1 -1 -.87 Door 7 .39 -.36 -1.63 -.30 -.37 Door 8 -4.57 -3.69 -1.63 -3.32 -3.67 Table 4.2: Leader expected values for each door selection in reward structure one Structure One Unobserved 5 20 Unlimited DOBSS 20% 7.5% 17.5% 12.5% COBRA(0,") 65% 65% 65% 70% COBRA(,") 57.5% 92.5% 72.5% 70% COBRA(C,") 92.5% 92.5% 87.5% 95% MAXIMIN 100% 100% 100% 100% Table 4.3: Percentage of times follower chose an expected strategy in reward structure one My first conclusion was that observational uncertainty is important. By accounting for this uncertainty, strategies are able to exploit human perceptions and make more appropriate use of resources. That is why COBRA(,") performs better than DOBSS in most conditions. In fact, examining Table 4.3 it can be seen that in the unobserved condition against COBRA(,"), human subjects played an expected strategy 57.5% of the time, such as door 3 or door 5 as seen in Table 4.2, while against DOBSS they played an expected strategy merely 20% of the time. However, based on MAXIMIN’s performance it is clear that getting followers to play expected strategies is not the only component. 76 As mentioned earlier, there is a trade-o in this match with expected strategies and leader ex- pected values. MAXIMIN is too loose with its resources making all responses expected strategies and thus the benefits begin to diminish because resources are spread too thin. It is important to utilize resources eciently and not squander them unnecessarily. By more accurately modeling human responses the new strategies of COBRA(;) are better able to utilize resources to guard against these responses and thus achieve a higher average expected value, one that is closer to the expected value they expect to receive based on their expected strategies. Of course when obser- vation is high it is assumed that observational uncertainty is low and the strategies are adjusted accordingly. Thus I find that utilizing a strategy that exploits human anchoring bias, but does not squander resources, provides the benefits I am seeking. My second conclusion was that addressing bounded rationality is important when dealing with human adversaries. I reach this conclusion due to COBRA(0,")’s superior performance. In fact, examining Table 4.3 it is clear that under all observation conditions the assumptions made by DOBSS are a poor model of human choices. Against COBRA(0,") on the other hand, which addresses bounded rationality by utilizing the concept of "-optimal responses, human subjects consistently play expected strategies 65-70% of the time under all observation conditions. While this improvement in expected strategy match comes at the cost of lower expected value (Ta- ble 4.2) the overall results indicate that the trade-o leads to an overall better performance of COBRA(0,"). This is a clear indication of the benefits that can be obtained by addressing bounded rational- ity. In fact, it can specifically be seen in the 20 and unlimited observation conditions for reward structure one and two that addressing bounded rationality provides improvements over DOBSS. Indeed, it is necessary to address bounded rationality when dealing with humans [Simon, 1956]. 77 Many times their choices can be guided by their cognitive limitations and thus it is necessary to robustly guard against a spectrum of possible choices rather than optimize against a single optimal choice [Simon, 1956; Rubinstein, 1998; Simon, 1969; Camerer, 2003]. By optimizing against the perfectly rational choice DOBSS may make poor use of its resources when dealing with human adversaries. Even COBRA(0,") is not a perfect model of human behavior, however, it is at least a step in the right direction since it is able to obtain expected strategies more often than algorithms without bounded rationality. I defer the explanation of the final key observation to Section 4.4.4. However, I point out that due to the poor performance of COBRA(,") in reward structure three compared to DOBSS I ran further experiments exploring dierent values that are separate from the experiments I will present in Section 4.4.4. In these additional experiments, using a strategy with = :5 I found through experimentation with 40 new subjects that in the unlimited observation condition, COBRA(,") obtains an average expected value of -1.35 outperforming DOBSS with an average expected value of -1.5. In Section 4.4.4, in addition to =:5 for reward structure three, I present 3 alternate choices for in each reward structure for the unlimited observation condition on top of the three choices presented in these original experiments. 4.4.4 Handling Observational Uncertainty Given the significant impact of on these results, this section provides further analysis of the choice of on performance. Given that human choice under uncertainty remains a key area of research in psychology [Fox and Clemen, 2005; Fox and Rottenstreich, 2003; Kahneman and Tversky, 1972; Koehler and James, 2009; Starmer, 2000; Tversky and Koehler, 1994], it is di- cult to provide a definitive answer; however, I provide a solid initial grounding and heuristics for 78 choosing. I focus on the two extreme observability cases — the unobserved observation case and the unlimited observation case — for this initial investigation. As explained before, since the choices made in the unobserved condition were made irrespec- tive of the strategy used (subjects did not have any information on the strategy being employed when they made their decision) it is possible to test all values in this case using the data col- lected for each reward structure. The results are presented in Figure 4.4 and demonstrate a clear increasing trend in all four reward structures. On the x-axis I vary the value of and on the y-axis I show the average expected value obtained for a particular value of given the choices made by the subjects in the unobserved condition. For example, looking at Figure 4.4(b), when = 0 the average expected value is -1.46 while when = 1 the average expected value is .7. These results suggest that in the absence of observations, humans do appear to be anchoring on the uniform distribution and thus = 1 is the optimal setting in all four reward structures. It follows that if the expectation is for humans not to take any observations of the leader strategy, exploiting their anchoring biases can be important. Determining an appropriate value for the unlimited observation condition is more dicult. Since the humans are given the strategy in advance under this condition and normally (with a few exceptions) changing also alters the strategy used, it is required to test each new value of with a new set of subjects. To avoid exhaustively testing a large number of values, I focused on testing three new values per reward structure in particular. Again, for these experiments I used the exact same setup described previously but the subjects only played 12 games instead of 14 (for the 12 new values) and each of these games were with unlimited observation. The new values of tested against human subjects were selected using two key criteria. Before discussing these criteria, I present Figure 4.5, which helps ground these criteria. I also 79 (a) Reward structure one (b) Reward structure two (c) Reward structure three (d) Reward structure four Figure 4.4: Unobserved condition - Expected average reward define a term which I will call strategy entropy as P n i=1 (p i log(p i )) where p i is the probability value shown on door i. Although this is the standard equation for entropy it diers in this setting as it is defined over the marginal distribution subjects are shown. This is in contrast to calculating the entropy of the mixed strategy that produced this marginal distribution. As explained previously in Section 4.4.2.3, in the unlimited observation condition, the strategies are presented to subjects as the marginal probability distribution of guards over the 8 doors, where the sum of the probabilities over all 8 doors will be 3. Specifically, for each door the subject is given a probability p and this is the probability that he will obtain his penalty (there will be a guard on the door) where 1 p is the probability he will obtain his reward (there will not be a guard on the door). Based on this definition, a higher strategy entropy represents a strategy where the probability value of each door is closer to:375 (300% divided evenly among 8 doors) and a lower strategy entropy represents a strategy where the probability value on each door is closer to 1 or 0 (more specifically the highest entropy possible is 4.245). Figure 4.5 shows, for each value of in each 80 reward structure, the strategy entropy produced from the corresponding mixed strategy. On the x-axis I list the value of and on the y-axis I list the strategy entropy obtained from the corre- sponding strategy. I highlight the values used in these additional experiments and the original experiments for the unlimited observation condition with a bold circle. More specifically, for each of these reward structures there are 5 highlighted values corresponding to the three new values selected for these additional experiments, the value used in COBRA(C,") in the original experiments, and finally the value used in COBRA(,") in the original experiments which was = 0 2 . (a) Reward structure one (b) Reward structure two (c) Reward structure three (d) Reward structure four Figure 4.5: Strategy entropy for varying values Given Figure 4.5, the two criteria for selecting values are as follows: i) The strategy en- tropy for the corresponding mixed strategy should not be low as this corresponds to deterministic strategies. Since humans are provided this strategy in advance, they would certainly exploit a 2 In reward structure three there is a sixth value that was used in the original experiments as an alternative value for COBRA(C,") in the unlimited observation condition to outperform DOBSS. 81 very high, i.e., low entropy. I selected strategies with entropy 1.19 or higher; and ii) The strat- egy entropy for the corresponding mixed strategy should be quantitatively dierent than other values already being tested. For example, for reward structure four, strategy entropy is constant over a range of values as the strategy is constant. Selecting two values from this range is not useful. The results of these new experiments along with the results from the original experiments in the unlimited observation condition can be seen in Figure 4.6. On the x-axis I present the dierent algorithms and dierent values of for COBRA(,") and on the y-axis I present the average expected value obtained by each of the corresponding strategies. For example in Figure 4.6(b), COBRA(.54,2.5) obtains an average expected value of -1.08. (a) Reward structure one (b) Reward structure two (c) Reward structure three (d) Reward structure four Figure 4.6: Average expected values for varying under the unlimited observation condition Figure 4.6 allows the following observations about setting in COBRA (,) for the unlim- ited observation condition to be made: 82 High values of lead to poor performance: Choosing a high , which leads to a lower strategy entropy performs poorly. Lower strategy entropy implies more determinism, which human followers exploit. In these experiments, strategy entropy below 2.4 appeared to degrade performance. Mid-range values of may lead to the best performance in reward structures where a fol- lower’s deviation from expectation is harmful to the leader: Previously it was assumed that = 0 would be the best setting for the unlimited observation condition, but these results are clearly contradictory to this. Choosing the lowest values of (near 0) leads to a high strategy entropy, but that does not provide the best outcome (in three of the four reward structures). One possible explanation for this result is that humans have diculty reason- ing about strategies that are overly complicated. If strategy entropy is too high then hu- mans may have diculty evaluating alternatives in order to find optimal or even-optimal strategies and this may cause them to deviate from expected strategies which can lead to a degradation in the leader’s expected value — as explained previously, these three reward structures were specifically designed such that deviation by the followers from expected strategies were harmful to the leader. In these experiments, a strategy entropy between 2.5 and 2.85 appears to be optimal. Lowest values of may perform optimally in zero-sum reward structures: In the fourth re- ward structure, or zero-sum reward structure, the value that gives the maximum strategy entropy (i.e., COBRA(0,2.5)) is optimal. This makes intuitive sense given that the optimal strategy in a zero-sum game is equivalent to a MAXIMIN strategy as shown by the observa- tions in Section 4.3. In a zero-sum game, any deviation from optimal play by the follower 83 leads to an expected value that is strictly higher for the leader. For reward structures one, two, and three however, deviations can lead to severe degradations in the leader’s expected value, making it more important to account for deviations from optimal play. The robust strategies in COBRA(;) are designed to combat these potentially detrimental deviations, which means in a zero-sum game, as previously explained, they are the least useful. How- ever, even in a zero-sum game it is still possible to exploit human anchoring biases in low observation conditions to obtain higher expected leader values as shown by these results. Based on these results we can see why COBRA(C,") is seen to outperform COBRA(,") in the unlimited observation condition as noted in Section 4.4.3.1. Also, these results have demon- strated that the optimal choice for in the unobserved observation condition appears to be = 1. In the unlimited observation condition these results have shown that, for general sum games, set- ting to a value that leads to a strategy between a mid-range to high-range entropy appears to be best. While a precise predictor for optimal remains an issue, particularly for other observation conditions, using a strategy entropy based technique for selecting in these conditions appears to be a promising approach. Given the analysis presented in Section 4.4.3.3 and these additional experiments, COBRA(,") and COBRA(C,"), with appropriately chosen and " values, appear to be the best performing among the presented algorithms. The performance of DOBSS in some of these experiments also illustrates the need for the novel approaches presented in this thesis for dealing with humans. For example, in Figure 4.3(a) it is clear that under high observation conditions DOBSS performs very poorly in comparison to other strategies. In fact, in this case DOBSS is seen performing even worse than simple baseline algorithms such as MAXIMIN. Indeed, with DOBSS having been deployed since August 2007 at Los Angeles International Airport (LAX) [Jain et al., 2010], 84 these results show that security at LAX could potentially be improved by incorporating these new methods for dealing with human adversaries. 4.4.5 Runtime Results For runtime results, in addition to the original 8-door game, I constructed a 10-door game with 10 3 = 120 leader actions, and 10 follower actions. To average the run-times over multiple in- stances, I created 19 additional reward structures for each of the 8-door and 10-door games. Fur- thermore, since the algorithms presented handle Bayesian games, I created 8 variations of each of the resulting 20 games to test scale-up in number of follower types. For the a priori probability distribution of follower types, I assume each follower type occurs with a 10% probability except the last which occurs with 1:10(n 1) probability where n is the number of follower types. For example, if there are 5 follower types, the first four types each occur with a 10% probability and the last type occurs with a 60% probability. Experiments were run using CPLEX 8.1 on an Intel(R) Xeon(TM) CPU 3.20GHz processor with 2 GB RDRAM. In Figure 4.7, I summarize the runtime results for the Bayesian game using DOBSS, COBRA(0,"), COBRA(,0), COBRA(,") and MAXIMIN. I include one graph for the 8-door results and one for the 10 door results. For COBRA(,") I set " = 2:5. For both COBRA(,0) and COBRA(,") I varied the value of to show the impact on solution speed. I include = :25 and = :75 in the graph, denoted by COBRA(.25,2.5)/COBRA(.75,2.5) for COBRA(,") and COBRA(.25,0)/COBRA(.75,0) for COBRA(,0) respectively. The x-axis in Figure 4.7 varies the number of follower types from 1 to 8. The y-axis of the graph shows the runtime of each algo- rithm in seconds. All experiments that were not concluded in 20 minutes (1200 seconds) were cut o. 85 As expected, MAXIMIN is the fastest among the algorithms with a maximum runtime of 0:054 seconds on average in the 10-door case. Not anticipated was the approximately equivalent runtime of DOBSS and COBRA(0,") and even more surprising were the significant speedups of COBRA(,") and COBRA(,0) over DOBSS and COBRA(0,") depending on the value of. As shown in Figure 4.7 as increases, the runtime of COBRA(,") and COBRA(,0) decreases. For example, in the 10-door 8 follower type case when = :25 COBRA(,") is unable to reach a solution within 1200 seconds on average, however, when is increased to :75, COBRA(,") is able to find a solution in 327.5 seconds on average. In fact, excluding MAXIMIN, every strategy except COBRA(,") with = :75 and COBRA(,0) reached the maximum runtime in the 10-door 8 follower type domain. This speedup could be attributed to the branch-and- bound methods used to find solutions to these MILPs. Since COBRA(,") and COBRA(,0) distribute some of their weight to the uniform distribution it decreases the number of branch- and-bound nodes necessary to achieve a solution by decreasing the branch space. This set of results particularly support this theory. These results demonstrate that COBRA(;) does not incur significant runtime costs for the proposed enhancements to deal with bounded rationality and observational uncertainty and in fact may even provide runtime improvements over DOBSS. Figure 4.7: Comparing runtimes 86 Chapter 5: MATCH Algorithm In Chapter 4 I demonstrated the value of addressing bounded rationality and observational un- certainty in COBRA(;). Since the inception of COBRA(;), dierent models have been proposed for addressing the bounded rationality of human adversaries [Pita et al., 2010; Yang et al., 2011]; however, in security game settings, the approach known as BRQR [Yang et al., 2011] presented in Section 2.5 has emerged as the leading approach under unlimited (i.e., per- fect) observation conditions. While COBRA(0;) avoids the dicult task of predicting the human-response function by instead taking an approach based on robust optimization [Aghassi and Bertsimas, 2006] that protects against worst-case deviations within an-bound, its robustness feature does not account for any deviations beyond the-bound. Thus, for deviations larger than, the defender (leader) can still suer arbitrarily large degradations in her expected value. BRQR is a model that attempts to model all potential deviations under perfect observation, which may be why it has been shown to outperform COBRA(0;). To that end, I significantly modify the approach proposed in COBRA(0;) based on the stan- dard worst-case assumption of robust optimization and, instead, bound the defender’s loss for a 87 potential deviation by the human attacker based on the degree of the deviation from the expected- value-maximizing strategy. This is done in an eort to also address all potential deviations as in BRQR, but still avoid the complex task of modeling human decision making by relying on a robust approach. This new algorithm, MATCH, provides three key benefits: (i) it provides significant runtime benefits over BRQR; (ii) it strongly couples the adversary’s and defender’s performance, robustly guarding against potential deviations by human adversaries and avoiding situations where minor deviations (i.e., deviations that result in minor losses in expected value) by the adversary may result in large losses for the defender; (iii) it avoids the dilemma of creating an accurate opponent model. I refer to this new type of optimization as graduated optimization and show in Section 5.1 that it lies within a space of robustness between MAXIMIN and the standard game-theoretic optimal solution. To evaluate the advantages of this new approach, I make the most comprehensive investigation to date. Specifically, I examine 104 security settings where I take the four recommended security settings from Yang et al. [2011] and I also intelligently select 100 additional payo structures, which I will describe in detail in Section 5.2. Furthermore, in Section 5.2 I defend why I believe future experimental setups should follow this particular experimental setup as a guideline. I test the 104 security game settings against 363 human subjects playing 8823 games in total to compare the performance of MATCH against BRQR. The results reveal that MATCH performs as well as or better than BRQR against human adversaries in over 90% of the settings tested and in Section 5.3 I give an analysis of these results. While MATCH provides a number of benefits under unlimited observation, it is not designed to address limited observation conditions as in COBRA(, ). Furthermore, MATCH is not as robust to minor deviations, which may be more likely, as COBRA(,) is. In the future it may 88 be beneficial to combine the two unique robust optimization ideas and the concept of anchoring biases into a single unified algorithm designed to be more robust for all observation conditions. 5.1 MATCH Algorithm The key concept behind the MATCH algorithm is the new idea of graduated robust optimization. Whereas standard robust optimization robustly guards against a worst-case outcome within some error bound, MATCH assumes an expected-value-maximizing outcome on behalf of the attacker, but constrains the impact of deviations depending on the magnitude of the deviation. Specifically, the defender’s loss for a potential deviation by the attacker is bounded based on the distance of that deviation from the expected-value-maximizing strategy. MATCH is specifically designed for security games, which I previously defined in Chapter 2 Section 2.6 and uses the same notation. However, in MATCH the following additional restrictions are made on the payos of the defender and attacker for covered and uncovered targets: U u : T! (1; 0) (5.1) U c : T! (0;1) (5.2) U u : T! (0;1) (5.3) U c : T! (1; 0) (5.4) This maintains the restrictions of U and U respectively and is more in line with real-world security domains where an attack creates a negative impact for the defender and a prevented attack a negative impact for the attacker. Additionally, I define2 [0;1) and restrict the attacker 89 to choosing a pure strategy i 2 . The justification for only considering attacker pure strategies is the same as previously described for DOBSS in Chapter 2 Section 2.4. In the MATCH MILP the defender’s goal is to maximize her expected value which I represent as V and the defender solves the following: max V s:t: X i2T c i = K (5.5) c i 2 [0::: 1] 8 i2 T (5.6) = arg max i 2 U (c; i ) (5.7) V U (c; ) (5.8) (U (c; ) U (c; ˆ ) V U (c; ˆ ) 8 ˆ 2 ; c ˆ < 1 (5.9) Constraints 5.5 and 5.6 ensure that the defender utilizes all her resources and that no target has more than 1 resource assigned to it. Constraint 5.7 ensures that the attacker chooses a target that maximizes his expected value. Constraint 5.8 ensures that the defender obtains the corresponding expected value (V) to the attacker’s optimal strategy. Constraint 5.9 is the most crucial portion of the formulation. The left portion calculates the attacker loss in expected value for a deviation from the optimal strategy. The right hand side constrains the defender loss in expected value for this deviation by the attacker to be no more than a factor of times the loss the attacker receives. For example, if the defender does not want to lose any more than twice what the attacker loses for a potential deviation, then we can set = 2. If the defender does not want to lose any more than 90 half what the attacker loses, then we can set = :5. This provides a direct trade-o between the defender’s maximum expected value for the attacker’s optimal strategy and additional protection on potential weaknesses for deviations. MATCH addresses important issues in both COBRA and BRQR. A fundamental property of MATCH is that it does not rely on some complex non-linear non-convex optimization problem (e.g., as in BRQR or other approaches such as RPT Yang et al. [2011]). Indeed, its power is in its perceived simplicity, which not only means it is simple to implement, but it is orders of mag- nitude faster than its competitors including BRQR. In addition, similar to BRQR, it allows for a more gradual defense against deviations as opposed to a hard cuto point as in the-rationality approach of COBRA. However, MATCH avoids the challenge of creating an accurate opponent model of human decision making by relying instead on a form of robust optimization. Nonethe- less, MATCH still faces one crucial consideration, which is a trade-o between robustness and defender expected value. The key dierence between MATCH and BRQR is that BRQR attempts to model human decision making; but if this model is inaccurate, the defender’s performance suers. MATCH in contrast bypasses modeling of the human-decision-making process; it instead directly focuses on how much maximum expected value a defender is willing to trade o to protect against the human attacker’s potential deviations from the rational strategy. While the-parameter can be adjusted, in the experimental sections I will consistently keep = 1 and show that even with this flat setting without any tuning, MATCH outperforms BRQR with careful tuning of. Furthermore, the performance of MATCH might also be enhanced with alternative-settings, however, finding an appropriate procedure for estimating the-parameter is left for future work. Proposition 1. If = 0, then MATCH maximizes the worst-case outcome for the defender. 91 Proof. If = 0 it follows that Constraint 5.9 becomes V U (c; ˆ ) 8 ˆ 2 ; c ˆ < 1. If c ˆ < 18 ˆ 2 , by definition Constraints 5.8 and 5.9 maximize the leader’s minimum expected value since V U (c; )8 2 . If9 c ˆ = 1 then I will show that U (c; ˆ ) U (c; 0 )8 0 2 ; c 0 < 1, guaranteeing that U (c; ˆ ) is the best worst-case bound by definition of a security game since c ˆ cannot be increased further. If there exists more than one target that is fully covered (i.e., c ˆ = 1) it is only necessary to consider the target that gives the defender the least expected value among them. Consider (c; ) an optimal solution for MATCH with = 0. Let c ˆ = 1 and assume9 0 2 : U (c; 0 ) < U (c; ˆ ); c 0 < 1. It follows from Constraint 5.9 that V U (c; 0 ) (here it may be the case that 0 = ). By definition of a security game, the defender’s expected value could be improved by increasing the value of c 0 and this could be accomplished by directly trading probability from c ˆ to c 0 at least until U (c; ˆ ) = U (c; 0 ), a contradiction since (c; ) is an optimal solution. Proposition 2. If is suciently large, then a coverage vector and an attacker strategy (c; ) that is optimal for the MATCH MILP corresponds to at least one SSE of the game 1 . Proof. As stated in Definition 1, in order for (c; ) to be a SSE they must meet three defined criteria: (i) the leader plays a best response, (ii) the follower plays a best response, (iii) the follower breaks ties optimally for the leader. If is suciently large then Constraint 5.9 is now trivially satisfied for any target not in the attack set ((c)), eectively removing it. For targets in the attack set, I will show that in any optimal solution the MATCH MILP will force the attacker (follower) to break ties in the defender’s (leader’s) favor. Then I will show that the MATCH MILP forces the defender and attacker to play mutually best responses. 1 In this case Constraint 5.9 is eectively removed making the MATCH MILP exactly equivalent to an existing MILP known as ERASER which has previously been shown to compute optimal solutions that correspond to at least one SSE of the game[Kiekintveld et al., 2009] 92 Consider a solution with 2 (c); , t where t is as defined in Section 2.6. It follows that U (c; ) = U (c; t ), U (c; ) U (c; t ), and V U (c; ). For target t , Constraint 5.9 now becomes 0 V U (c; t ). If U (c; t ) > V then the constraint is satisfied, however, the defender could do better if the attacker deviates from to t . By definition of a security game, the defender can induce this deviation by removing an arbitrarily small coverage from c t and adding it to c , thus inducing the favorable SSE by selecting a strategy arbitrarily close to the equilibrium [von Stengel and Zamir, 2004]. This change would create a coverage vector c c with an attack set that would only contain t and the defender would receive utility V U (c; t ). To verify, by removing an arbitrarily small coverage probability from c t , it has forced U (c; t ) > U (c; t) 8t2 T; t , t , including any other target previously in the attack set since no coverage probability was removed from other targets. In the opposite direction, if 9 ˆ 2(c); U (c; ˆ )< U (c; ) then Constraint 5.9 is no longer satisfied. In order to satisfy this constraint the MILP would force the attacker to choose the less favorable target ( ˆ ) and by the same logic the favorable outcome could be induced, removing the unfavorable outcome from the set. So far I have shown that the MATCH MILP forces the attacker to break ties in the defender’s favor. It remains to be shown that the MATCH MILP forces the defender and attacker to play best responses as defined in a SSE. In MATCH, Constraint 5.8 defines the defender’s expected value (V) contingent on the target attacked ( ). The constraint places an upper bound of U (c; ) on V. Since the objective maximizes V for any optimal solution V = U (c; ). This also implies that c is maximal, given for any optimal solution, since V is maximized. In a similar way, Constraint 5.7 forces the attacker to select a target that maximizes his expected value given a coverage vector 93 c. Taken together, the objective and Constraints 5.7, 5.8 and 5.9 (for tie breaking) imply that (c; ) are mutual best-responses in any optimal solution. While I have shown that (c; ) are mutual best responses, it remains to be shown that c cor- responds to a mixed strategy x2 X for the defender and that the attacker’s Stackelberg strategy F can be constructed. As previously stated, it has been shown that the coverage vector c2 C can be converted into an equivalent mixed strategy x2 X [Kiekintveld et al., 2009]. While c corresponds to a mixed strategy for the defender, is an incomplete description of the attacker’s Stackelberg strategy F ; it does not specify choices for any coverage other than c. I will show that the constraints of the MILP imply the existence of a function F extending such that c and F satisfy the criteria of a SSE. While I have shown that c and are mutual best-responses for an optimal MILP solution, it remains to describe the attacker’s behavior for any other feasible coverage vectors c 0 , c. Let 0 = t 2 (c 0 ) be a target in the attack set for c 0 with maximal expected value for the defender. By construction, 0 is feasible in the MILP and satisfies criteria 2 for a SSE. Based on the selection of 0 = t 2 (c 0 ), remember that Constraint 5.9 will ensure that U (c 0 ; t) = U (c 0 ; t ) 8 t2 (c 0 ), satisfying criteria 3 for a SSE. Since (c 0 ; 0 ) is a feasible solution in the MILP, U (c 0 ; 0 ) U (c; ) since (c; ) is optimal for the MILP. Let F be a function constructed using this method for every possible c 0 , c. c is a best-response to F since U (c 0 ; 0 ) U (c; ), satisfying the first criteria of a SSE. 94 5.2 Experiment Purpose, Design, and Results 5.2.1 Purpose of this Study I sought to investigate the performance of BRQR and DOBSS against MATCH under perfect observation. Here, performance is measured by the average expected value obtained by a security force against the decisions of human adversaries. In these experiments I only examine one crucial variable of real-world domains, which is the reward structure, while the other variables are held constant. For BRQR I fixed =:75 and for MATCH I fixed = 1:0. While I discuss estimating appropriate-settings in Section 5.4, my choice of =:75 was inspired by the original estimate made by Yang et al. [2011]. I chose = 1 to constrain the defender losses to be no worse than the attacker losses for deviations. In Section 5.2.2.2 I will describe how the reward structures were chosen. The goal of MATCH was to increase the performance of a security force against human adver- saries by addressing the bounded rationality that humans may exhibit under perfect observation conditions. To that end, experiments were set up where human subjects would play as followers (adversaries) against each strategy under dierent reward structures given perfect observation. It is not possible to prove optimality against human adversaries who may deviate from the expected optimal responses and thus I rely on empirical validation through experimentation. In addition to examining the performance of BRQR with MATCH, I also examine the runtime performance of MATCH against BRQR. 95 5.2.2 Experimental Design Similar to the experimental design in Chapter 4, these experiments are inspired by the real-world security domain at LAX [Pita et al., 2008] described previously. Here, three guards – jointly acting as the defender – guard eight gates, and an individual human subject acts as a single attacker who will choose one of the eight possible gates. Again, the 8 gates model the 8 terminals found at LAX. The interface can be seen in Figure 5.1 Figure 5.1: Game Interface Each subject was modeled as a single follower type defined by the reward structure they were given. In order to simulate the Stackelberg setting, subjects were presented with the following information before they chose a gate: (i) the subject’s reward and penalty for each gate; (ii) the defender strategy (i.e., the probability distribution of the guards over the 8 gates); (iii) the guard’s reward and penalty for each gate. This can be seen in Figure 5.1. Again, I use this Stackelberg framework because in real-world scenarios an attacker can conduct extensive surveillance of his potential target and the corresponding defensive strategy before choosing to attack, which would allow him to learn this information. 96 In each game instance, the guards would choose 3 gates based on their strategy and the sub- ject’s goal was to choose the gate that would maximize expected value given the defender’s strat- egy. Each of the 8 gates would have a dierent reward and penalty associated with it for both the subjects as well as the guards. For instance, as shown in Figure 5.1, gate 4 has a reward of 7 and a penalty of -8 for the subject and a reward of 7 and a penalty of -1 for the guard. If a subject chose a gate that a guard was guarding, the subject would incur the subject penalty for that gate and the guard would receive the guard reward for that gate, else vice-versa. Going back to the previous example, if the subject chose gate 4 and a guard was guarding that gate then the subject would receive -8 and the guard would receive 7. This setup led to a Stackelberg game with 8 3 = 56 leader actions, and 8 follower actions. 5.2.2.1 Participants These experiments were run in Amazon Mechanical Turk with the requirement that workers were from the United States. Outside of the qualification that the workers were from the United States, the workers were anonymous except for an identifying worker number. This worker number was tracked in order to insure a worker could not participate in more than one of the following experiments. There are three dierent experimental evaluations that were conducted in this study. In the first experimental evaluation there were 69 unique (i.e., did not participate in any of the other experiments or more than once in this experiment) anonymous Amazon Mechanical Turk workers from the United States. In the second experimental evaluation there were 253 unique anonymous Amazon Mechanical Turk workers from the United States. In the third experimental evaluation there were 41 unique anonymous Amazon Mechanical Turk workers from the United States. 97 5.2.2.2 Reward Structure Given this experimental setup, I ran experiments in two sets of reward structures. I first explored four reward structures proposed by Yang et al. [2011], which were chosen to be the most rep- resentative of the entire payo structure space for security games based on metrics proposed by Yang et al. [2011]. These reward structures can be found in the Appendix under Section B, Tables B.5-B.8. However, I found that in these particular reward structures the strategies produced by BRQR and MATCH were highly similar. Thus, to more fully compare the performance of BRQR and MATCH, for the second set of experiments I systematically chose 100 new reward structures based on covariant games in GAMUT [Nudelman et al., 2004]. For Covariant games, GAMUT creates a game for a given number of players with payos distributed normally(0,1) with covariance r2 [1; 1], which determines a covariance between player rewards. Specifically, when r =1 the game is zero-sum between players and when r = 1 the game is perfectly cooperative. I slightly modified the code in GAMUT to restrict the resulting payos to meet the criteria of a security game. That is, the payos fall within a specified range of [-10,-1] for penalties and [1,10] for rewards. I chose covariant games to generate payo structures because they are naturally able to capture the adversarial nature of security games. To select the 100 new reward structures I first generated 1000 reward structures with the r parameter ranging from -.9 to 0 by .1 increments (i.e., 100 games for each setting of r). I chose 0 as the boundary because, in an adversarial security setting, it does not make sense that the reward and penalty vectors would be positively correlated. In an attempt to examine how dierent the MATCH and BRQR algorithms are overall, for each structure I computed the 1-norm distance between MATCH and BRQR where, as I explained previously, I set =:75 and = 1. 98 (a) 1-Norm (b) Selected Structures Figure 5.2: 1-Norm Scatter Plots In Figure 5.2(a) I present the scatter plot for these 1000 reward structures. For readability, on the x-axis I display the setting of r from 0 to:9 as 0 to 9 (i.e., on the x-axis 3 represents -.3). On the y-axis I display the 1-norm value. These scatter plots show that there is a wide range of possibilities for the dierence between MATCH and BRQR strategies. While extreme dierences don’t occur as frequently, they do not appear to be completely rare. Given these 1000 structures, I attempted to select 100 reward structures that would best cover the possible space based on the 1-norm values. I present the 100 reward structures selected within the scatter plot in Figure 5.2(b) and the actual reward structures can be found in the Appendix under Section B, Tables B.9-B.108. I believe this procedure for selecting reward structures is superior to previous procedures [Pita et al., 2010; Yang et al., 2011] for three critical reasons: i) by examining the 1-norm distance I can explore a spectrum of reward structures where the strategies produced are most dierent (high 1-norm) to where they are most similar (low 1-norm); ii) by utilizing covariant games, I can control the correlation between player rewards to ensure rewards and penalties are not positively correlated where previous experiments have ignored this crucial issue [Yang et al., 2011; Pita et al., 2010]; and iii) previous experimental results may give a distorted view of the 99 overall performance of an algorithm compared to other algorithms since they look at such a narrow portion of the entire security game space (i.e., 4 to 10 potential settings versus over 100). 5.2.2.3 Experimental Procedure All of these experiments were run in Amazon Mechanical Turk and participants were paid a base amount of US $1.50 for participating. For the first experiment, I tested the mixed strategies generated by DOBSS, BRQR and MATCH in the first set of reward structures (i.e., the four structures proposed by Yang et al. [2011]). I used the original estimate made by Yang et al. [2011] (i.e., = :76) and = 1 to constrain the defender’s losses to be no worse than the attacker’s losses for MATCH. Each of the 12 game settings (four reward structures and three algorithms) were played by 40 subjects (i.e., in total there were 480 total trials). The 12 potential game settings were broken into two separate groups of 6 game settings. The first group of 6 game settings consisted of all combinations of game settings for reward structures 5 and 6 (i.e., 2 reward structures and three algorithms). Similarly, the second group consisted of the same for reward structures 7 and 8. All experiments were first conducted for group 1 with 40 subjects and then group 2 with 40 dierent subjects. The following procedure was used for both groups. First, to ensure that subjects were not choosing gates arbitrarily, I introduced two obvious games where a gate with the highest reward and lowest penalty possible had the lowest probability (5%) of being covered. These games and corresponding strategies can be found in the Appendix under Section G.2.1, Tables G.1 and G.2. If subjects did not choose this gate in these two games their results were removed from the final set. This lead to less than 40 subjects in the final data sets presented, which accounts for the number of participants presented in Section 5.2.2.1. 100 In order to mitigate the order eect on subject responses, a total of 8 dierent orderings of the 8 game settings (6 original settings and 2 dummy game settings) were generated using Latin Square design. Every ordering contained each of the 8 game settings exactly once, and each game setting appeared exactly once in each of the 8 positions across all 8 orderings. The order played by each subject was drawn uniformly randomly from the 8 possible orderings. Before beginning the game, subjects were given a brief tutorial about how to play to ensure that they understood the general game play. This tutorial can be found in the Appendix under Section G.3. After beginning, subjects were given an unlimited amount of time to make a decision on which gate to choose for each game. As stated earlier, subjects were paid a base amount of US $1.50 for participating in this study. However, to further motivate the subjects, they were allowed to earn additional bonus money based on their performance in the game. For each reward point earned or lost, a subject would receive or lose an additional US $0.15 from their net money. Before playing, subjects were informed that only 5 of their games would be selected from the 8 games played to determine the actual bonus payment. Since subjects were not aware which games would be chosen, they would have incentive to perform the best they could in each game and they were given immediate feedback at the end of each game. Subjects were paid their final earnings, but were not required to pay money back if their net points were less than zero. For the second set of experiments, I tested the mixed strategies generated by BRQR and MATCH in the second set of reward structures (i.e., the 100 structures chosen using Gamut). I chose to omit DOBSS in these experiments since it has been previously shown that DOBSS performs poorly against human adversaries [Yang et al., 2011], and the previous experiment along with the experiments with COBRA also demonstrate this result. This lead to a total of 200 101 possible game settings (100 reward structures and two algorithms). The exact same procedure described for the first experiment was used for this experiment, except the 200 potential game settings were grouped dierently. To avoid boredom in the subjects, I limited the number of games they would have to play by separating the reward structures into the following groups played by the number of subjects indicated: (i) Reward structures 9-13 (I start at 9 to account for the previous 9 reward structures) [30 participants (25 after removal)], (ii) Reward structures 14-25 [40 participants (33 after re- moval)], (iii) Reward structures 26-40 [40 participants (37 after removal)], (iv) Reward structures 41-57 [45 participants (40 after removal)], (v) Reward structures 58-74 [40 participants (37 after removal)], (vi) Reward structures 75-91 [45 participants (42 after removal)], and (vii) Reward structures 92-108 [40 participants (39 after removal)]. Subjects would play against both BRQR and MATCH for a given group of reward structures. The only dierence between this experi- mental setup and that described for the first experiment was the number of games played and the corresponding Latin Square. For example, subjects who played reward structures 14-25 played a total of 26 game settings (24 game settings from the reward structures and 2 from the dummy games), which lead to 26 dierent orderings of the 26 game settings. As before, every ordering contained each of the 26 game settings exactly once, and each game setting appeared exactly once in each of the 26 positions across all 26 orderings. The order played by each subject was drawn uniformly randomly from the 26 possible orderings. For the third set of experiments I tested the mixed strategies generated by BRQR and MATCH. Again, I chose to omit DOBSS for the reasons explained in the second experiment. The same procedure was used as in the previous two experiments. I leave discussion on which game settings were tested for Section 5.4 since it relies on the results of the second experiment. 102 In the following, I will first present the results for the reward structures proposed by Yang et al. [2011], then the results for the newly selected structures, and finally I will give an analysis of these results. I evaluate the statistical significance for all results using the bootstrap-t method [Wilcox, 2003] used by Yang et al. [2011] previously, which is described in the Appendix under Section A. 5.2.3 Results for Original Structures I present the results of these experiments in Figure 5.3. In total, 36 subjects played against reward structures 5 and 6 while 33 played against reward structures 7 and 8 (after removal of subjects who failed the obvious games). In Figure 5.3, the y-axis represents the average defender expected value over all the choices made by each individual subject against a particular strategy. For example, examining reward structure 5 in Figure 5.3, we see that the defender received -0.29 on average against human subjects if using the strategy generated by BRQR. Figure 5.3: Original reward structures In reward structures 5 and 8 MATCH is statistically significantly better than BRQR and DOBSS (p:028 and p:004 respectively). In reward structures 6 and 7 BRQR and MATCH are not statistically significantly better than each other (p = :15 and p = :392 respectively), 103 but both are statistically significantly better than DOBSS (p = 0). However, in general across all 4 payo structures, MATCH and BRQR create highly similar strategies (i.e., the probability dierence on any particular gate is relatively low [ 12.6%]). Regardless, even with the sim- ilar strategies, MATCH still outperforms BRQR with statistical significance in two of the four reward structures and does at least as well in the other two. Based on these results I can con- clude that there exists conditions where MATCH is the superior algorithm to BRQR and once again accounting for human adversaries is crucial since both algorithms significantly outperform DOBSS. 5.2.4 Results for New Reward Structures While the results from the first experiment were promising, I wanted to take the most extensive look to date at a large space of potential security game settings, examining 100 potential reward structures compared to 10 or less in previous experiments [Yang et al., 2011; Pita et al., 2010]. I present an overview of the results from these experiments in Table 5.1. Here, I show the number of settings where MATCH won with statistical significance, both strategies were approximately equivalent (i.e., neither strategy won with statistical significance), and BRQR won with statistical significance. In 42 of the 100 reward structures MATCH outperformed BRQR with statistical significance and in an additional 52 of the 100 reward structures MATCH did at least as well as BRQR given that neither strategy won with statistical significance. These results combined with the previous experiment show MATCH performing at least as well as or outperforming BRQR in 98 out of 104 potential security settings. In Section 5.3 I will give further analysis of these results. 104 MATCH Draw BRQR =:05 42 52 6 Table 5.1: Overview of Results MATCH Draw BRQR TOTAL Rejected 40 40 3 83 Not Rejected 2 12 3 17 Table 5.2: Pearson Chi-squared Results 5.3 Analysis The QR model is a well-established solution concept and so an important question to address is whether BRQR was actually an accurate model of human decision making in these security settings. To determine whether BRQR is accurate, in each reward structure I run a Pearson’s chi-squared goodness of fit test [Greenwood and Nikulin, 1996] on the predicted distribution of attacker choices against the observed attacker choices for the subjects. I present the results in Table 5.2. In the first three columns I denote reward structures where MATCH won with sig- nificance, neither strategy won with significance, and BRQR won with significance. In the last column I give the overall result for all 100 reward structures. In the rows I denote whether Pear- son’s chi-squared goodness of fit test rejected the null hypothesis that the observed distribution of choices could have been drawn from the expected distribution of choices ( =:05). My first observation is that in 83% of all reward structures tested the model proposed by BRQR did not fit the data observed. This is a significant number and suggests that perhaps BRQR is not a good model of human decision making in security games. However, it is possible that this result is due to a poor estimation of the-parameter for these particular security settings and so in Section 5.4 I will re-estimate the-parameter based on the observed data and run additional experiments for a key subset of the reward structures. Even so, in real-world security settings it 105 may be even more dicult to appropriately estimate since data can often be sparse or noisy, and the problem instances can be much larger and more complex. The fact that, for this set of results, BRQR does not provide a good fit for the data observed in general is one potential explanation for why MATCH is outperforming BRQR in the majority of the security settings. BRQR attempts to exploit an assumed model of the human attacker and if the attacker deviates from that model in a significant way it can severely impact the performance of the defender. For instance, if BRQR assumes that an attacker is not likely to attack a certain target it will provide minimal coverage for that target. If, however, a large number of attackers choose this target, contrary to the prediction, it can have severe eects on the defender’s average expected value. An inaccurate model of human decision making can lead to severe consequences in security domains. This is one of the key advantages of MATCH since it does not assume any decision-making model on behalf of the attacker and specifically bounds the impact of such potential deviations. My second observation is that of the 17 structures where BRQR was potentially a good fit of human decision making, it only outperformed MATCH in 3. Thus, of the six cases where BRQR won with statistical significance, only three of the cases can potentially be justified by accurate opponent modeling. These results show that even a decent model of human decision making may not be sucient enough and an approach based on robust optimization is potentially a strong alternative to modeling human-decision-making processes. It may be necessary to have a highly accurate model of human decision making before it becomes sucient in security settings. A second important question is, given the space of possible reward structures, is it possible to determine when MATCH or BRQR will likely perform better. In Figure 5.4, I present where the results for the 100 structures appear in the scatter plot of 1000 structures. In the cases where 106 neither MATCH nor BRQR won with statistical significance I present which strategy had the higher average defender expected value. While this does not imply that the strategy is better in these cases, this was done to see if it would provide some insight into what portion of the payo structure space MATCH or BRQR performed better. Here, it is evident that the results are diverse within the potential space of reward structures and so without further investigation I cannot de- termine where BRQR will likely be better than MATCH overall. However, these results support my earlier argument that when doing experimental analysis in security settings, examining few potential settings can give misleading results. For example, for a small sample of potential reward structures chosen throughout the potential space, it is possible I could have chosen only structures where BRQR outperformed MATCH, which would be contrary to what I have been able to show in my results. Thus, I have rightly raised the standard for future experimental investigations of new potential algorithms. Figure 5.4: Scatter Plot of Results 107 5.4 -Re-estimation As suggested in my analysis, in order to confirm whether BRQR is actually a poor predictor of human decision making I am required to examine BRQR with appropriately estimated - values. To focus my analysis I selected three groups of five reward structures from the 42 reward structures where MATCH outperformed BRQR with statistical significance as follows: (i) the five reward structures where BRQR and MATCH had the most significant strategy dierence averaged over the 1-norm, 2-norm, infinite-norm, and KL distances; (ii) the five reward structures where BRQR and MATCH had the least dierence in average expected value; and (iii) the five reward structures where BRQR and MATCH had the highest dierence in average expected value. For these experiments I will refer to these as reward structures 1 through 15. To re-estimate the - parameter I used the Maximum Likelihood Estimation procedure proposed by Yang et al. [2011] using the data from the previous experiments in each of the 15 reward structures yielding 15 new -values. In Table 5.3 I present the new-estimates along with the 1-norm distance between the strategy produced by the original-setting ( =:75) and the re-estimated-setting. My first observation is that the value of is largely dependent on the reward structure im- plying that for each potential security domain the defender would be required to make a new estimate. As stated previously, estimating can already be dicult in real-world settings where data may be sparse or noisy. This problem is further exacerbated since I have now shown that data cannot likely be pooled from dierent settings. This is an additional advantage of the approach in MATCH since the level of robustness is not dependent on the reward structure. That is, once the level of robustness has been decided (i.e., -setting) it is consistent across all reward structures where a-setting is not equivalently accurate for all reward structures. 108 Structure: 1 2 3 4 5 6 7 8 : .18 .71 .25 1.14 .01 1.39 1.09 .67 1-norm: .376 .012 .384 .177 3.10 .221 .110 .050 Structure: 9 10 11 12 13 14 15 : .84 .48 .15 .43 .55 .23 .42 1-norm .018 .069 .424 .396 .127 .244 .356 Table 5.3: New-estimates My second observation is that the actual impact of altering the-parameter varies significantly depending on the reward structure. For example, in structure 10 I vary from .75 to .48 and see only a 1-norm dierence of .069 while in structure 12 I reduce to .43 and see a 1-norm dierence of .396 (i.e., there is a 40% dierence of probability across targets in one case and only a 7% dierence across targets in another case). This once again demonstrates the diculty in appropriately estimating since minor changes can lead to significant dierences in some reward structures. I had 41 subjects play against both MATCH ( = 1) and BRQR with newly estimated - values in all 15 reward structures using the procedure outlined in Section 5.2.2.3. I present the results in Figure 5.5. In these results, MATCH remained statistically significantly better in 8 of the 15 reward structures (structures 1-5, 11, 13, and 15) and neither strategy was statistically significantly better in the remaining 7. Additionally, I ran Pearson’s chi-squared goodness of fit test and found that, even after re-estimation, the null hypothesis that the observed choice distribution could have been drawn from the predicted choice distribution was rejected in all 15 cases. Thus, even if I obtain tailored estimates, MATCH continues to perform as well as or outperform BRQR. 109 Figure 5.5: Re-estimated Reward Structures 5.5 Runtime Results In Figure 5.6 I present runtime results for BRQR versus MATCH. In Figure 5.5, the number of resources are fixed at 10 and on the x-axis I vary the number of targets from 10 to 50. On the y-axis I present the runtime in seconds averaged over 20 randomly generated payo structures. In Figure 5.5, the number of targets are fixed at 30 and on the x-axis I vary the number of resources from 2 to 20. Once again on the y-axis I present the runtime in seconds averaged over 20 randomly generated payo structures. Experiments were run using CPLEX 8.1 on an Intel(R) Core(TM) i5-2450M CPU 2.50GHz processor with 6 GB RAM. These results show that MATCH provides orders of magnitude speedup over BRQR further demonstrating the benefits of such an approach. The reason for this runtime improvement is that BRQR requires the solution to a non-linear and non-convex objective function in its most gen- eral form. In fact, because of the complexity of the objective function, BRQR is only a heuristic solution for solving the objective. MATCH on the other hand is a mixed-integer linear program, which can be solved with standard packages. These results demonstrate that MATCH signifi- cantly improves runtime costs for the proposed enhancements to deal with bounded rationality. 110 (a) Increasing Targets (b) Increasing Resources Figure 5.6: Runtime results Based on these experimental results I have demonstrated five fundamental contributions of the MATCH algorithm: (i) I develop an approach to addressing human adversaries based on robust optimization rather than relying on finding appropriate models of human decision-making; (ii) I introduce a systematic way to generate meaningful reward structures based on covariant games where previous work has simply generated completely random reward structures; (iii) I make the most comprehensive evaluation to date involving 363 human subjects playing 8823 games in 104 security game settings; (iv) I demonstrate that MATCH performs as well as or better than BRQR in over 94% of the security settings tested (42 of 104 settings with statistical significance); and (v) I demonstrate the significant runtime benefits of MATCH over BRQR. These results demon- strate the potential benefits of using an approach based on robust optimization (e.g., MATCH or COBRA) over previous algorithms that rely on creating more ecient models of human-decision making. 111 Chapter 6: Security Circumvention Games There are three strict assumptions I aimed to address in this thesis. Namely, that human adver- saries are perfectly rational, that they perfectly observe the leader’s strategy, and that their action space is tractable. In Chapters 4 and 5 I addressed issues surrounding the assumptions of per- fect rationality and perfect observability. However, the action space assumption remains to be addressed. To motivate the importance of addressing this assumption, I will use a real-world sce- nario faced by the United States Transportation Security Administration (TSA) that lead to my new modeling approach and an application that is under evaluation [Pita et al., 2011]. The TSA is tasked with protecting the nation’s transportation systems [TSA, 2012]. These systems are often large in scale and require many personnel and security activities to protect them. One set of systems in particular is the over 400 airports [TSA, 2012]. These airports serve approximately 60 million aircraft annually [NATCA, 2012]. To protect this large transportation network, the TSA employs approximately 48,000 Transportation Security Ocers [TSA, 2012]. These Security Ocers are responsible for implementing security activities at each individual airport in order to provide security for the transportation network. While many people are aware of common security activities, such as individual passenger screening, this is just one of many security layers TSA personnel implement to help prevent 112 potential threats [TSA, 2012]. These layers can involve hundreds of heterogeneous security ac- tivities executed by limited TSA personnel leading to a complex resource allocation challenge. Unfortunately, the TSA cannot possibly run every security activity all the time and thus must de- cide how to appropriately allocate its resources among the layers of security to protect against a number of potential threats. To aid the TSA in scheduling resources in a risk-based manner, I take a multi-agent game-theoretic approach. From the perspective of the underlying game-theoretic model, a crucial dierence of my novel approach and previous approaches is this: I allow for both heterogeneous security activities and threats whereas the security games presented previously are only able to consider homogeneous security activities and threats, leading to a new game model called “Security Circumvention Games” (SCGs). In conjunction with TSA subject matter experts, I developed a software system, Game- theoretic Unpredictable and Randomly Deployed Security (GUARDS), that utilizes a Stackelberg framework to aid in protecting the airport transportation network. However, the TSA’s security challenge raises two new critical issues. The first issue is in appropriately modeling the TSA’s security challenges in order to achieve the best security policies (mixed strategy). Due to the com- plex nature of TSA’s security challenges, traditional models of security games [Korzhyk et al., 2011] are no longer appropriate models. Specifically, the TSA’s domain has the following addi- tional features beyond traditional security games: (i) heterogeneous security activities for each potential target; (ii) heterogeneous threats for each potential target; (iii) unique security activities for individual airports. The second issue is in eciently solving the model I developed where, because I consider a national deployment, a special-purpose solver may not be appropriate. In fact, previous solution techniques [Jain et al., 2010; Kiekintveld et al., 2009; Jain et al., 2011b; Yin and Tambe, 2012] for traditional security games are no longer directly applicable. 113 To address these issues, I developed both a new formal model of security games and tech- niques to solve this class of games. To appropriately model the TSA’s security challenges I created a novel game-theoretic model, which is referred to as Security Circumvention Games (SCGs), and cast the TSA’s challenges within this model. In the creation of SCGs I provide the following contributions: (i) the ability for defenders to guard targets with more than one type of security activity (heterogeneous activities), and (ii) the ability for attackers to choose threats designed to circumvent specific security activities. Given my new model, I designed an ecient solution technique in which I create a compact representation of SCGs. This allows the applica- tion to avoid using a tailored Stackelberg solver and instead utilize a general purpose Stackelberg solver to compute solutions eciently. 6.1 TSA Security Challenges I now describe in detail the two major issues in potentially deploying game-theoretic random- ization for airport security on a national scale, including modeling and computational challenges along with my solutions to them. 6.1.1 Modeling the TSA Resource Allocation Challenges While I am motivated by an existing model of security games [Korzhyk et al., 2011], there are three critical aspects of the new TSA domain that raise new challenges. First, the defender now reasons over heterogeneous security activities for each potential area within an airport 1 . For example, airports have ticketing areas, waiting areas, and cargo holding areas. Within each of these areas, the TSA has a number of security activities to choose from such as perimeter patrols, 1 Due to the nature of the TSA’s security challenge, I will refer to targets in the TSA’s domain as areas henceforth. 114 screening cargo, screening employees and many others. Second, given the multiple possible security activities, the defender may allocate more than one resource per area (i.e., areas are no longer covered or uncovered). Finally, the defender now considers an adversary who can execute heterogeneous attacks on an area. The TSA must reason about a large number of potential threats in each area such as chemical weapons, active shooters, and bombs. The key challenge is then how to allocate limited TSA security resources to specific activities in particular areas, taking into account an attacker’s response. To address this challenge it is necessary to create a more expressive model than outlined in security games; one that is able to reason over the numerous areas, security activities, and threats within an individual airport. We refer to this new class of security games as Security Circumvention Games (SCGs). SCGs are more expressive than traditional security games and thus can represent both traditional security games and the games I consider for the TSA. In SCGs, the TSA must choose some combination of security activities to execute within each area and the attacker must reason over both which area to attack and which method of attack to execute based on the defender’s strategy. At this time I elaborate on the defender’s and attacker’s possible strategies. 6.1.1.1 Defender Strategies I still denote the defender by , and the set of defender’s pure strategies by i 2 . The TSA is able to execute a variety of security activities, which I denote by S =fs 1 ;:::; s jSj g. Each security activity has two components. The first is the type of activity it represents, and the second is the area where the activity is performed. I denote the set of areas by A =fa 1 ;:::; a jAj g. 115 The defender still has R =fr 1 ;:::; r jRj g resources available and thus a pure strategy, i 2 , is a subset of security activities from S with size equal tojRj. The TSA’s task is to consider how to allocate these resources among security activities in order to provide the optimal protection to their potential areas. For example, if there are three security activities, S =fs 1 ; s 2 ; s 3 g and two resources available, one possible pure strategy for the defender is to assign these two resources to s 1 and s 3 . Given that the number of possible combinations ofjRj security activities at an airport can be on the order of 10 13 or greater for the TSA, I develop a compact representation of the possible strategies that I present in Section 6.1.2. The defender’s mixed strategies 2 are the possible probability distributions over . Similar to previous work, a mixed strategy (randomized solution) is typically the optimal strategy. 6.1.1.2 Attacker Actions Defending an area against terrorist attacks is complicated by the diversity of the potential threats. For example, an attacker may try to use a vehicle borne explosive device, an active shooter, a suitcase bomb, and many others in any given area. Not all methods of attack would make sense in all areas. For example, using a vehicle borne explosive device in the checked baggage screening area in some airport configurations would not be a viable method of attack. I still denote the attacker by , and the set of pure strategies for the attacker is still given by j 2 . Each pure strategy for the attacker corresponds to selecting a single area a i 2 A to attack, and a specific mode of attack. However, given the complexity of the TSA’s security challenge and that human adversaries are capable of continually evolving new threats, enumerating all the possible threats may not be practical if even possible. To address the action space assumption described 116 previously and avoid this diculty, I developed a novel way to represent threats for the TSA’s domain that I describe in Section 6.1.2.1. 6.1.2 Compact Representation for Eciency While I have developed a model that appropriately captures the TSA’s security challenge, one issue with this model is that both the attacker and defender strategy spaces grow combinatorially as the number of defender security activities increases. Also, listing such a large number of potential threats would lead to extreme memory and runtime ineciencies. Furthermore, existing solution techniques that have been developed for security games [Jain et al., 2010; Kiekintveld et al., 2009; Jain et al., 2011b; Yin and Tambe, 2012] are not directly applicable to Security Circumvention Games (SCGs). With this in mind, I looked at an alternate approach to finding optimal solutions eciently. Specifically, I looked at representing threats in a more intelligent manner and creating a compact representation for the defender strategy space. By utilizing both of these techniques, I achieved large reductions in runtime. I utilized DOBSS [Paruchuri et al., 2008] to solve my compact rep- resentation and avoided creating a tailored algorithm for each specific airport. At this time I will explain both how I model threats and how I achieve a compact representation of the defender’s full strategy space. 6.1.2.1 Threat Modeling for TSA While it is important to reason over all the security activities that are available to an individual airport, enumerating all of the large number of potential threats they face can lead to severe memory and runtime ineciencies. Thus, the problem I face is how to model attack methods 117 in a way that limits the number of threats the game-theoretic model needs to reason over, but appropriately captures both an attacker’s capabilities and his goals. In particular, I automatically generate attack methods for the adversary that capture two key goals: (i) an attacker wants to avoid the security activities that are in place, and (ii) an attacker wants to cause maximal damage with minimum cost. In order to achieve these goals an intelligent adversary will observe security over time and design his attack method based on his observations. The attacker’s plan will be designed to avoid security activities that he believes will be in place. I will refer to this as circumventing security activities. For example, imagine there is a single area with three security activities such as passenger screening, luggage screening, and perimeter patrol. In this example, the TSA only has one resource available and thus can only execute one of these activities at a time. While passenger screening may have the highest probability of success, if TSA personnel never screen luggage or patrol the perimeter, the adversary can choose an attack path that avoids passenger screening such as utilizing a suitcase bomb or an attack from the perimeter. On the defender side, dedicating more resources to security activities in an area increases the security aorded to that area. However, even with more resources, the TSA wants to avoid being predictable since attackers can exploit this predictability; avoiding the security activities they know will be in place. Thus, I needed to represent threats in a way that accounts for the attacker’s ability to observe security in advance and avoid specific security activities, but still represents the benefit of dedicating more resources. A na¨ ıve approach is to represent only a single threat per area and decrease the likelihood of success for that threat as more security activities are put in place. This captures the increase in security for additional security activities, however, it does not account for the attacker’s ability to 118 circumvent security activities. With this method you would simply choose security activities in the order of their relative success making it predictable and exploitable. The alternative that I chose is to create a list of potential threats that circumvent dierent combinations of specific security activities. By basing threats on circumventing particular com- binations of security activities, I avoid the issue of enumerating all the possible potential threats. Instead the threats are automatically created based on the security activities in an area. However, I also incorporate a cost to the attacker for circumventing more activities to capture the idea of causing maximal damage at minimal cost. Each individual activity has a specific circumvention cost associated with it and more activities circumvented leads to a higher circumvention cost. This cost reflects the additional dicultly of executing an attack against increased security. This diculty could be due to requiring additional resources, time and other factors for executing an attack. Since attackers can now actively circumvent specific security activities, randomization becomes a key factor in the solutions that are produced because any deterministic strategies can be circumvented. 6.1.2.2 Compact Representation I introduce a compact representation that exploits similarities in defender security activities to reduce the number of strategies that must be enumerated and considered when finding an optimal solution to SCGs. First, I identify security activities that provide coverage to the same areas, and have the same circumvention costs (i.e., have identical properties). Let i 2 K represent the sets of security activities that can be grouped together because they have identical properties. Now, instead of reasoning over individual security activities, I reason about groups of identical security 119 activities i 2 K. A strategy i 2 is represented by the number of resources assigned to each set of identical security activities i . To illustrate this new representation, I provide a concrete example of the full representation versus the compact representation in Tables 6.1 and 6.2. In this example there are 4 security activities and 2 resources. Here, s 1 and s 2 have identical circumvention costs and aect a 1 while s 3 and s 4 have identical circumvention costs and aect a 2 . Table 6.1 presents the full represen- tation with corresponding payos and Table 6.2 represents the compact form of the same where 1 represents the group s 1 and s 2 and 2 represents the group s 3 and s 4 . In both tables, each row represents a single pure strategy for the defender and each column the same for the attacker. Notice in Table 6.1 each strategy i 2 is represented by the exact security activities being executed while in Table 6.2 it is only which set i 2 K each resource has been allocated to. The key to the compact representation is that each of the security activities from a set i 2 K will have the same eect on the payos. Therefore, it is optimal for the defender to distribute probability uniformly at random across all security activities within a set i , so that all security activities are chosen with equal probability in the solution. Given that the defender strategy uni- formly distributes resources among all security activities s j 2 i , it does not matter which specific security activities the attacker chooses to circumvent from the set i . For any given number of se- curity activities circumvented, the expected payo to the attacker is identical regardless of which specific activities within the set are chosen. This is because security activities are being selected uniformly at random within the set i . Therefore, I can use a similar compact representation for the attacker strategy space as for the defender, reasoning only over the aggregate number of security activities of each type rather than specific security activities. 120 Given this, I only need to know how many security activities are selected from each set in order to compute the expected payos for each player in the compact representation. For example, examining the second row and second column of Table 6.2 we see that the reward to the defender is -2 and the reward to the attacker is 0. In this case, the defender strategy is to assign 1 resource to activities in 1 and 1 resource to activities in 2 . Given that she is uniformly distributing these resources, it follows that she will execute s 1 half of the time and s 2 the other half. On the attacker side, we know that the attacker is circumventing one security activity from the set 1 . If he circumvents either s 1 or s 2 he will only succeed half of the time. Thus, half of the time the defender receives 4 and the other half -8 for an expectation of -2 (i.e., 4:5 + (8):5). I compute the attacker’s reward in the same manner. a 1 :; a 1 : s 1 a 1 : s 2 a 2 :; a 2 : s 3 a 2 : s 4 s 1 ; s 2 2, -1 4, -3 4, -3 -20, 10 -17, 7 -17, 7 s 1 ; s 3 2, -1 -8, 3 4, -3 5, -5 -17, 7 8, -8 s 1 ; s 4 2, -1 -8, 3 4, -3 5, -5 8, -8 -17, 7 s 2 ; s 3 2, -1 4, -3 -8, 3 5, -5 -17, 7 8, -8 s 2 ; s 4 2, -1 4, -3 -8, 3 5, -5 8, -8 -17, 7 s 3 ; s 4 -10, 5 -8, 3 -8, 3 5, -5 8, -8 8, -8 Table 6.1: Example payos for sample game. a 1 :; a 1 : 1 a 2 :; a 2 : 2 1 ; 1 2, -1 4, -3 -20, 10 -17, 7 1 ; 2 2, -1 -2, 0 5, -5 -4.5, -5 2 ; 2 -10, 5 -8, 3 5, -5 8, -8 Table 6.2: Example compact version of sample game. Given this compact representation for both the defender and attacker, I can compute an opti- mal mixed strategy of assigning resources over K. Once I have this mixed strategy, I will need to determine an actual strategy for TSA personnel to execute by sampling one of the possible strategies from the mixed strategy I have determined for my compact representation (e.g., one 121 sample may be 2 2 ). Once sampled, I will know exactly how many resources are available to each set i 2 K. Given this resource assignment, I can then sample security activities by selecting m uniformly at random where m is the number of resources assigned to i 2 K. This specific set of security activities for each area under the current resource assignment is a full strategy for TSA personnel to execute. 6.2 Evaluation When evaluating a game-theoretic model like SCGs there are two important issues that are raised. The first issue is evaluating the value of the security policies generated against alternative ap- proaches. The second issue is with scalability and run-times. To be useful in practice, the ap- proach needs to be able to solve real world challenges. In the following sections I present each of these evaluations. 6.2.1 Security Policy Analysis For this analysis I examined the security policies generated by my game representation against two other possible solution strategies. The first strategy is a solution concept where resources are distributed uniformly among areas (uniformly random), an approach sometimes used in lieu of a game-theoretic approach. The second strategy uses my new representation, however, it does not allow attackers to circumvent security activities (SCGs without circumvention). That is, I allow the attacker only a single attack strategy per area and simply reduce the value of that strategy as the number of security activities increases. This is a simplified model of an attacker as mentioned 122 in Section 6.1.2.1. Finally, I included my new representation and allow an intelligent attacker to circumvent specific security activities when planning his mode of attack (SCGs). I generated 20 random game instances with 10 areas and 3 security activities per area. In each game instance the payo value of each area for both the defender and attacker are ran- domly selected from 1 to 50 and the circumvention costs are similarly selected from 1 to 5. I then calculated the optimal solution under the current solution strategy (i.e., uniformly random, SCGs without circumvention, and SCGs). After finding the optimal solution, I determined the expected value for each solution given the assumptions made in SCGs (i.e., attackers are allowed to circumvent specific security activities when planning their attack). For each game instance, I computed the optimal solution varying the number of resources available from 1 to 10 as seen on the x-axis of Figure 6.1. On the y-axis, I present the average expected value obtained by each solution strategy across all 20 game instances. In Figure 6.1 the results show that the uniform policy is outperformed by both game-theoretic approaches with the approach accounting for cir- cumvention strategies performing the best. In fact, an approach that accounts for circumvention strategies is the only one that was able to obtain a positive reward for the defender in the 20 randomly generated game instances and in the 10 resource case obtains a 200% improvement in reward over any other strategy. This shows the benefits of reasoning about an intelligent attacker who will research and exploit deterministic security activities. Figure 6.1: Policy Analysis: Increasing resources for 10 areas with 3 security activities per area 123 6.2.2 Runtime Analysis I present simulation results focusing on the computational eciency of my compact method versus the full representation. All experiments are run on a system with an Intel 2 GHz processor and 1 GB of RAM. I used a publicly available linear programming package called GLPK to solve optimization problems as specified in the original DOBSS procedure. For the compact version I use a slightly altered version of DOBSS that is designed specifically for eciency in the compact representation. The solver was allowed to use up to 700 MB of memory during the solution process. For larger game instances, solving the problem with the full representation runs out of memory and solutions cannot be found. In the results presented below I exclude results for cases where the full representation was not able to produce a result using the allotted memory. I also note that in all experiments both the solution found by the full representation and the solution found by the compact representation are optimal. To test the solution methods I generated random game instances by randomly selecting payo values from 1 to 50 and circumvention costs from 1 to 5 for each area. For each experiment I generated 20 random game instances and averaged the results (there is little variance in the run- times for dierent problem instances). I considered three dierent scenarios. The first scenario presents results for the case where there is an increasing number of areas, and each area has exactly 3 security activities associated with it. There are 5 resources available for the defender, and each security activity has identical properties (i.e., no security activity has a higher cost for circumvention or higher probability of success) for the area it is associated with. Given thejAj possible areas, for the full representation there are 3jAj 5 possible defender pure strategies and 8jAj possible attacker pure strategies. Thus, in the 10 area case there are 142,506 defender pure 124 strategies and 80 attacker pure strategies. Examining Figure 6.2 (a), I show the improvement in runtime of my compact representation over the full representation. For more than 4 areas, the full representation failed to achieve a solution within the memory bounds. For 4 areas, the compact representation runs much faster than the full representation, with a runtime of less than 1 second versus the 177 seconds required by the full representation. In fact, for 10 areas, the compact representation has an average runtime of approximately 1 second, which is still much faster than the full representation for only 4 areas. Even if the number of security activities associated with each area is a relatively small constant my compact representation provides substantial benefits. As the number of similar security activities associated with an area increases, this advantage grows. In the second scenario, I considered a situation where security activities are distributed ran- domly across possible areas. The total number of security activities is set similarly to the previous experiment, in that the total number of security activities is three times the number of areas. How- ever, I randomly assigned security activities to areas (with each area having at least one security activity) so the number is no longer uniform across areas. Once again the defender has 5 re- sources available and security activities have identical properties within an area. It follows that in the full representation, the number of defender pure strategies and attacker pure strategies are identical to the previous scenario. However, the number of strategies in the compact represen- tation for both the defender and attacker may vary. Looking at Figure 6.2 (b), the results show similar benefits for the compact representation in this case as in the previous experiment with a uniform distribution of activities. In the final scenario, I considered a situation in which there are 10 areas to protect, each area has 3 identical security activities, and I increased the number of resources available to distribute 125 between these areas. Thus, in the full representation, assuming there arejRj resources available, the defender has 30 jRj possible pure strategies and the attacker has 80 possible pure strategies. In Figure 6.3, I increase the number of resources available along the x-axis and show the time to compute a solution in seconds on the y-axis. The full representation is unable to compute a solution for more than 4 resources under these conditions within the allotted memory. On the other hand, the compact representation is able to arrive at a solution for 10 available resources in less than 30 seconds. These results show the benefits of my compact representation in terms of eciency. Even so, this last result shows that even the compact representation can face runtime diculties for larger numbers of resources. Given the scale and complexity of airports, finding improved computational methods for this compact representation remains an area for future work. In general though, SCGs provide an ecient framework for addressing more complex security scenarios than traditional security games. (a) Three activities per area (b) Random activities per area Figure 6.2: X-axis: Areas, Y-axis: Runtime Figure 6.3: Runtime: Increasing resources for 10 areas with 3 security activities per area 126 Chapter 7: Conclusions 7.1 Summary Stackelberg games have become crucial in many multiagent applications, and particularly for se- curity applications [Brown et al., 2006; Paruchuri et al., 2008; Jain et al., 2010; Shieh et al., 2012; Pita et al., 2011]; for instance, these games are applied for security scheduling at the Los Angeles International Airport, the Federal Air Marshal Service, and the United States Coast Guard [Jain et al., 2010; Shieh et al., 2012]. In such applications, automated Stackelberg solvers may create an optimal leader strategy. Unfortunately, the standard game-theoretic solution techniques makes three critical assumptions that may not apply in the real-world against a human adversary: Rationality Assumption: One of the strict assumptions made by standard game-theoretic approaches based on Stackelberg games is that the adversary is perfectly rational. That is, the adversary maximizes his expected value based on the information available. Observability Assumption: In standard Stackelberg game models it is assumed that the adversary perfectly observes the security force’s strategy. 127 Action Space Assumption: Game-theoretic models require defined potential actions on be- half of all the players involved in order to compute an optimal mixed strategy for the secu- rity force. If these assumptions fail to hold against a human adversary, then it may lead to a severely under-performing strategy when the human adversary deviates from the optimal strategy. In fact, human decisions are guided by their bounded rationality [Simon, 1956, 1969] as opposed to expected-value-maximizing rationality, which may cause them to deviate from their expected op- timal strategy. Furthermore, in real-world settings, humans may have limited observability of the security personnel’s strategy, giving them a false impression of that strategy. Given this limited information, humans can be biased in their decision making, even deviating from the optimal re- sponse of the observed strategy Fox and Rottenstreich [2003]; See et al. [2006a]. Beyond issues of rationality and observability, real-world security scenarios are complex in nature with a large potential action space. Additionally, human adversaries are adaptable and creative allowing them to constantly evolve new attack strategies. Simply enumerating such a large action space may not be practical and, given current solution methods, may not be solvable under reasonable memory or time limitations. This thesis helped address these three critical assumptions and provides three key contribu- tions: COBRA: COBRA includes two new key ideas for addressing human adversaries: (i) hu- man anchoring biases drawn from support theory, and (ii) robust approaches for MILPs to address human imprecision. To the best of my knowledge, the eectiveness of each of these key ideas against human adversaries had not been explored in the context of Stackelberg 128 games. Furthermore, it was unclear how eective the combination of these ideas, being brought together from dierent fields, would be against humans. MATCH: MATCH extends the robustness approach of COBRA allowing for a more grad- ual defense against potential deviations as opposed to a hard cut-o. This new type of graduated robust optimization avoids the complex task of modeling human decision mak- ing while still accounting for dierent degrees of potential deviations. Security Circumvention Games: SCGs are a novel game-theoretic model, which help address the vast and continually evolving action space of real-world security domains. In creating this model I provide the following contributions: (i) the ability for defenders to guard areas with more than one type of security activity (heterogeneous activities), and (ii) the ability for attackers to choose threats designed to mitigate specific security activities. Furthermore, I designed an ecient solution technique for reasoning over SCGs where I rely on creating a compact representation of each game instance and solving it using a general purpose Stackelberg solver. The experimental evaluations of COBRA and MATCH validate the usefulness of these ap- proaches against human adversaries compared to both game-theoretically optimal algorithms and an approach known as BRQR, which has previously been shown to be the best performing against human adversaries under perfect observation [Yang et al., 2011]. For COBRA, these results are based on 4 reward structures, in 4 dierent observability conditions, involving 218 human sub- jects playing 2960 games in total. For MATCH, these results are based on 104 reward structures, under perfect observation, involving 363 human subjects playing 8823 games in total. Further- more, MATCH provides significant runtime benefits over BRQR since it exploits the unique 129 structure of security games [Korzhyk et al., 2011]. The empirical evaluation of SCGs reveals both the benefits to solution quality against intelligent human adversaries and the runtime ben- efits of such an approach when addressing the vast action space of human adversaries, which may not be practical to enumerate in the standard way. These new approaches for addressing human adversaries provide a significant contribution in transitioning game-theoretic approaches to real-world security domains where security forces will face human adversaries. 7.2 Future Work In the future the application of game-theoretic approaches will continue to expand to new se- curity domains. For example, security agencies may use game-theoretic approaches to create randomized patrols to protect forests from illegal deforestation or to create randomized patrols in national parks to protect animals from illegal poachers. Each new domain will bring with it new challenges and new complexities. However, the common element in each new security domain will be the human adversaries that security forces will likely face. This thesis has shown the value of addressing the standard critical assumptions of game- theoretic approaches. However, in each new domain the type of adversary may change, creating a need for more evolved approaches for addressing human adversaries. In addressing the challenges of human adversaries I envision building on the basic insights developed in my thesis, which includes the value of robustly guarding against potential deviations and intelligently representing the underlying action space of security games. In the short run, the robustness approaches presented could be enhanced by exploiting ad- ditional human biases and further developing the graduated robustness ideas of MATCH. For 130 example, MATCH could be extended to incorporate COBRA’s-robustness for minor deviations, address human anchoring biases under limited observation and to vary the degree of-robustness based on the magnitude of deviations. That is, instead of being equivalently robust to large devia- tions as small deviations, an approach could be created that allows the defender to sacrifice more as the magnitude of the deviation increases. In the longer run, it will be crucial to address the dierent adversary types that exist in dif- ferent security domains. Depending on the type of adversary faced, the human-decision-making process may vary. In the future it will be important to explore dierent psychological profiles and how they aect human-decision-making processes. By examining these crucial psychological profiles it may be possible to create superior algorithms designed specifically for certain criminal types and security domains. Additionally, in the future it will be important to explore the potential multi-attribute aspects of these security problems. MATCH is an algorithm that already indirectly explores the multi-attribute aspects of security problems by taking into consideration the impact a deviation has to the defender and not just the attacker. By doing so, MATCH indirectly accounts for the possibility that the attacker may place some value on the impact his deviation has to the defender. Such an attacker may be willing to sacrifice some of his expected value in order to significantly impact the defender’s expected value. In the future it would be beneficial to create a suite of potential algorithms that exploit dier- ent human-decision-making processes for specific psychological profiles, robustly guard against the likely deviations human adversaries will make, and potentially consider multiple attributes within security settings. Additionally, creating better mechanisms for testing and validating new approaches will become critically important. As these approaches continue to transition into real-world domains, appropriately testing and validating the models will become a key concern. 131 For instance, this may include evaluating things such as robustness to dierent types of uncer- tainty (e.g., observational uncertainty, payo uncertainty, or capability uncertainty), scalability, and deterrence value. For now, no formal methods exist for making comprehensive and reliable evaluations. 132 Bibliography M. Abrahms. What terrorists really want: Terrorist motives and counterterrorism strategy. Inter- national Security, 32(4):78–105, 2008. M. Aghassi and D. Bertsimas. Robust game theory. Math. Program., 107(1):231–273, 2006. N. Agmon, G.A. Kaminka, and S. Kraus. Multi-robot adversarial patrolling: Facing a full- knowledge opponent. Journal of Artificial Intelligence Research, 42:887–916, 2011. G. Allison and P. Zelikow. Essence of Decision: Explaining the Cuban Missile Crisis. Longman, 1999. B. An, D. Kempe, C. Kiekintveld, E. Shieh, S. Singh, M. Tambe, and Y . V orobeychik. Security games with limited surveillance. In AAAI, pages 1241–1248, 2012. C. Archibald and Y . Shoham. Modeling billiards games. In AAMAS, pages 193–199, 2009. A. Azaria, Z. Rabinovich, S. Kraus, C. V . Goldman, and Y . Gal. Strategic advice provision in repeated human-agent interactions. In AAAI, pages 1522–1528, 2012. T. Basar and G. J. Olsder. Dynamic Noncooperative Game Theory. Academic Press, 1995. N. Basilico, N. Gatti, and F. Amigoni. Leader-follower strategies for robotic patrolling in envi- ronments with arbitrary topologies. In AAMAS, pages 500–503, 2009. M. Breton, A. Alj, and A. Haurie. Sequential stackelberg equilibria in two-person games. Opti- mization Theory and Applications, 59(1):71–94, 1988. G. Brown, M. Carlyle, J. Kline, and K. Wood. A two-sided optimization for theater ballistic missile defense. Operations Research, 53:263–275, 2005. G. Brown, M. Carlyle, J. Salmer´ on, and K. Wood. Defending critical infrastructure. Interfaces, 36(6):530–544, 2006. E. Brunner, U. Munzel, and M. L. Puri. Rank-score tests in factorial designs with repeated measures. Journal of Multivariate Analysis, 70(2):286–317, 1999. C. Camerer. Behavioral Game Theory: Experiments in Strategic Interaction. Princeton Press, 2003. V . Conitzer and T. Sandholm. Computing the optimal strategy to commit to. In ACM EC, pages 82–90, 2006. 133 C. de Melo, P. Carnevale, and J. Gratch. The eect of expression of anger and happiness in computer agents on negotiations with humans. In AAMAS, pages 937–944, 2011. DHS. United States Department of Homeland Security: Border Security. September 2012a. URL http://www.dhs.gov/border-security/. DHS. United States Department of Homeland Security: Transportation Security. September 2012b. URLhttp://www.dhs.gov/topic/transportation-security/. EPA. US Environmental Protection Agencey. September 2012. URLhttp://www.epa.gov/. C. R. Fox and R. T. Clemen. Subjective probability assessment in decision analysis: Partition de- pendence and bias toward the ignorance prior. Management Science, 51(9):1417–1432, 2005. C. R. Fox and Y . Rottenstreich. Partition priming in judgement under uncdertainty. Psychological Science, 14:195–200, 2003. M. Friedman. The use of ranks to avoid the assumption of normality implicit in the analysis of variance. Journal of the American Statistical Association, 32(100):675–701, 1937. N. Gatti. Game theoretical insights in strategic patrolling: Model and algorithm in normal-form. In ECAI, pages 403–407, 2008. P. Gill and J. K. Young. Comparing role-specific terrorist profiles. 2011. URL http://dx.doi.org/10.2139/ssrn.1782008. P. E. Greenwood and M. S. Nikulin. A Guide to Chi-squared Testing. John Wiley & Sons, Inc., 1996. GTI. Global tiger initiative ocial website. September 2012. URL http://www.globaltigerinitiative.org/. G. Haim, Y . Gal, and S. Kraus. A cultural sensitive agent for human-computer negotiation. In AAMAS, pages 451–458, 2012. J. C. Harsanyi and R. Selten. A generalized Nash solution for two-person bargaining games with incomplete information. Management Science, 18(5):80–106, 1972. M. Jain, J. Tsai, J. Pita, C. Kiekintveld, S. Rathi, M. Tambe, and F. Ord´ o˜ nez. Software assis- tants for randomized patrol planning for the LAX Airport Police and the Federal Air Marshals Service. Interfaces, 40:267–290, 2010. M. Jain, D. Korzhyk, O. Vanek, V . Conitzer, M. Pechoucek, and M. Tambe. A double oracle algorithm for zero-sum security games on graphs. In AAMAS, pages 327–334, 2011a. M. Jain, M. Tambe, and C. Kiekintveld. Quality-bounded solutions for finite Bayesian Stackel- berg games: Scaling up. In AAMAS, pages 997–1004, 2011b. M. P. Johnson, F. Fang, M. Tambe, and H. J. Albers. Patrol strategies to maximize pristine forest area. In AAAI, pages 37–41, 2012. 134 D. Kahneman and A. Tversky. Subjective probability: A judgement of representativeness. Cog- nitive Psychology, 3:430–454, 1972. D. Kahneman and A. Tversky. Prospect theory: An analysis of decision under risk. Econometrica, 47(2):263–292, 1979. G. L. Keeney and D. von Winterfeldt. Identifying and structuring the objectives of terrorists. Risk Analysis, 30(12):1803–1816, 2010. C. Kiekintveld, M. Jain, J. Tsai, J. Pita, M. Tambe, and F. Ord´ o˜ nez. Computing optimal random- ized resource allocations for massive security games. In AAMAS, pages 689–696, 2009. C. Kiekintveld, J. Marecki, and M. Tambe. Approximate method for infinite bayesian stackelberg games: Modeling distributional payo uncertainty. In AAMAS, pages 1005–1012, 2011. D. J. Koehler and G. James. Probability matching in choice under uncertainty: Intuition versus deliberation. Cognition, 113:123–127, 2009. D. Korzhyk, Z. Yin, C. Kiekintveld, V . Conitzer, and M. Tambe. Stackelberg vs. Nash in se- curity games: An extended investigation of interchangeability, equivalence, and uniqueness. volume 41, pages 297–327, 2011. LAPD. Ocial website of THE LOS ANGELES POLICE DEPARTMENT. September 2012. URLhttp://www.lapdonline.org/. LAWA. LAX - general description. September 2012. URL http://www.lawa.org/welcome LAX.aspx?id=40. G. Leitmann. On generalized Stackelberg strategies. Optimization Teory and Applications, 26 (4):637–643, 1978. K. Lye and J. M. Wing. Game strategies in network security. International Journal of Information Security, 4(1–2):71–86, 2005. S. Marsella, J. Gratch, and P. Petta. Computational models of emotion. In A blueprint for aective computing: A sourcebook and manual. Oxford: Oxford University Press, 2010. M. D. McCubbins, M. B. Turner, and N. Weller. The theory of minds within the theory of games. In International Conference on Artificial Intelligence, 2012. R. McKelvey and T. Palfrey. Quantal response equilibria for normal form games. Games and Economic Behavior, 10(1):6–38, 1995. A. R. Morral and B. A. Jackson. Understanding the role of deterrence in counterterorism security. RAND Corporation, OP-281-RC, 2009. URL http://www.rand.org/pubs/occasional papers/OP281/. NATCA. Air trac controllers. September 2012. URL http://www.natca.org/who we are.aspx?zone=Who We Are&pID=254#p254. 135 A. Nilim and L. E. Ghaoui. Robust control in markov decision problems with uncertain transition matrices. Operations Research, 53(5):780–798, 2005. Nobelprize.org. The sveriges riksbank prize in economic sciences in memory of alfred nobel 2000. September 2012. URL http://www.nobelprize.org/nobel prizes/economics/laureates/2000/. E. Nudelman, J. Wortman, Y . Shoham, and K. Leyton-Brown. Run the GAMUT: A comprehen- sive approach to evaluating game-theoretic algorithms. In AAMAS, pages 880–887, 2004. F. Ord´ o˜ nez and N. Stier-Moses. Robust wardrop equilibrium. In NET-COOP, volume 4465, pages 247–256, 2007. P. Paruchuri, J. P. Pearce, J. Marecki, M. Tambe, F. Ord´ o˜ nez, and S. Kraus. Playing games with security: An ecient exact algorithm for Bayesian Stackelberg games. In AAMAS, pages 895–902, 2008. J. Pita, M. Jain, C. Western, C. Portway, M. Tambe, F. Ord´ o˜ nez, S. Kraus, and P. Paruchuri. Deployed ARMOR protection: The application of a game theoretic model for security at the Los Angeles International Airport. In AAMAS, pages 125–132, 2008. J. Pita, M. Jain, M. Tambe, F. Ord´ o˜ nez, and S. Kraus. Robust solutions to stackelberg games: Addressing bounded rationality and limited observations in human cognition. Artificial Intel- ligence Journal, 174(15):1142–1171, 2010. J. Pita, C. Kiekintveld, M. Tambe, E. Steigerwald, and S. Cullen. GUARDS - game theoretic security allocation on a national scale. In AAMAS, pages 37–44, 2011. M. R. Pogrebin. About Criminals: A View of the Oenders’ World. SAGE, 2012. R. Porter, A. Ronen, Y . Shoham, and M. Tennenholtz. Mechanism design with execution uncer- tainty. In UAI, pages 414–421, 2002. L. Richardson. What Terrorist Want: Understanding the Enemy, Containing the Threat. Random House, 2006. H. Roso and R. John. Decision analysis by proxy for the rational terrorist. In QRASA, pages 25–32, 2009. A. Rubinstein. Modeling Bounded Rationality. MIT Press, 1998. T. Sandler and D. G. Arce M. Terrorism and game theory. Simulation and Gaming, 34(3): 319–337, 2003. K. E. See, C. R. Fox, and Y . Rottenstreich. Between ignorance and truth: Partition dependence and learning in judgment under uncertainty. Journal of Experimental Psychology: Learning, Memory, and Cognition, 32:1385–1402, 2006a. K. E. See, C. R. Fox, and Y . Rottenstreich. Between ignorance and truth: Partition dependence and learning in judgment under uncertainty. Journal of Experimental Psychology: Learning, Memory, and Cognition, 32:1385–1402, 2006b. 136 R. Selten. Evolutionary stability in extensive two-person games - correction and further develop- ment. Mathematical Social Sciences, 16:223–266, 1988. E. Shieh, B. An, R. Yang, M. Tambe, C. Baldwin, J. DiRenzo, B. Maule, and G. Meyer. PRO- TECT: An application of computational game theory for the security of the ports of the United States. In AAAI, pages 2173–2179, 2012. H. Simon. Rational choice and the structure of the environment. Psychological Review, 63(2): 129–138, 1956. H. Simon. Sciences of the Artificial. MIT Press, 1969. C. Starmer. Developments in non-expected utility theory: The hunt for a descriptive theory of choice under risk. Journal of Economic Literature, XXXVIII:332–382, 2000. D. Stevens, T. Hamilton, M. Schaer, D. Dunham-Scott, J. J. Medby, E. W. Chan, J. Gibson, M. Eisman, R. Mesic, C. T. Kelly Jr, J. Kim, T. LaTourrette, and J. Riley. Implementing security improvement options at Los Angeles International Airport. RAND Corporation, 2006. URLhttp://www.rand.org/pubs/documented briefings/2006/RAND DB499-1.pdf. S. Tijs. Nash equilibria for noncooperative n-person games in normal form. SIAM Review, 23(2): 225–237, 1981. TSA. Transportation Security Administration — U.S. Department of Homeland Security. September 2012. URLhttp://www.tsa.gov/. J. Tsai, Z. Yin, J. Kwak, D. Kempe, C. Kiekintveld, and M. Tambe. Urban security: Game- theoretic resource allocation in networked physical domains. In AAAI, pages 881–886, 2010. A. Tversky and D. Koehler. Support theory: A nonextensional representation of subjective prob- ability. Psychological Review, 101(4):547–567, 1994. J. von Neumann. Zur theorie der gesellschaftsspiele. Mathematische Annalen, 100:295–320, 1927. H. von Stackelberg. Marktform und Gleichgewicht. Springer, 1934. B. von Stengel and S. Zamir. Leadership with commitment to mixed strategies. In CDAM Re- search Report LSE-CDAM-2004-01, London School of Economics, 2004. R. Wilcox. How many discoveries have been lost by ignoring modern statistical methods? Amer- ican Psychologist, 53(3):300–314, 1998. R. Wilcox. Applying Contemporary Statistical Techniques. Academic Press, 2003. R. Wilcox. Introduction to Robust Estimation and Hypothesis Testing. Academic Press, 2005. J. R. Wright and K. Leyton-Brown. Beyond equilibrium: Predicting human behavior in normal- form games. In AAAI, pages 901–907, 2010. R. Yang, C. Kiekintveld, F. Ord´ o˜ nez, M. Tambe, and R. John. Improving resource allocation strategy against human adversaries in security games. In IJCAI, pages 458–464, 2011. 137 Z. Yin and M. Tambe. A unified method for handling discrete and continuous uncertainty in Bayesian Stackelberg games. In AAMAS, pages 855–862, 2012. Z. Yin, M. Jain, M. Tambe, and F. Ord´ o˜ nez. Risk-averse strategies for security games with execution and observational uncertainty. In AAAI, pages 758–763, 2012a. Z. Yin, A. Jiang, M. Johnson, M. Tambe, C. Kiekintveld, K. Leyton-Brown, T. Sandholm, and J. Sullivan. TRUSTS: Scheduling randomized patrols for fare inspection in transit systems. In IAAI, 2012b. K. K. Yuen. The two-sample trimmed t for unequal population variances. Biometrika, 61:165– 170, 1974. J. Zhuang and V . Bier. Secrecy and deception at equilibrium, with applications to anti-terrorism resource allocation. Defence and Peace Economics, 22(1):43–61, 2011. 138 Appendix A: Statistical Significance Tests As noted earlier, I was unable to use classical tests such as the T-test to judge statistical signifi- cance and hence dierent types of tests were run on my data. Additionally, because the experi- mental setup for COBRA was dierent than that used for MATCH, I use separate statistical tests for the COBRA data and the MATCH data. I will first review the statistical tests used for COBRA and why I chose these tests and then I will do the same for MATCH. A.1 Statistical Significance Tests for COBRA Among the reward structures used to test COBRA, reward structure four was treated separately as it is a zero-sum game. I explain the tests I used and the reasons these were chosen in the fol- lowing. Reward structure one and reward structure two are structures with higher penalties to the leader when the follower deviates from strong Stackelberg equilibrium (SSE) assumptions. Given this similarity in reward structures one and two and the similarity of the results they produced I first ran a two-way Friedman test for repeated observations Friedman [1937] in the unobserved observation condition and a two-way Friedman test (not for repeated observations) in the 5, 20, and unlimited observation conditions between them. A two-way test is a test that examines two variables within an experiment. In my case the two variables are reward structure and algorithm. I ran these tests to be certain that reward structure had an impact on the results produced for re- ward structures one and two. If reward structure did not have an impact on the results produced the data sets can be combined into one large data set. In the unobserved observation condition, the conclusion based on this test was that the reward structures did have an impact on the results, however, in the 5, 20, and unlimited observation conditions I found that they did not. Since reward structure did not influence the results in these cases, the data sets from reward structures one and two were combined into one large data set for the 5, 20, and unlimited observation cases. This arrangement left reward structure three separated by itself (since reward structure four was kept separate as a zero-sum game), but when combining reward structure three with one and two, I find that it had an impact on performance and hence separation was justified. The null hypothesis in all of my tests is that the results produced between any two algorithms are actually identical. This hypothesis is rejected with a p-value of .05 or less. Given this setup I had the following statistical tests. First, the unobserved case was treated separately from other cases. In the unobserved case, subjects were only given the reward structure and were asked to make a choice. Consequently, the choice they made would be made irrespective of the actual strategy employed. Therefore I was able to take the choice made by an individual subject and employ it across DOBSS, COBRA(0,"), 139 COBRA(,"), COBRA(C,"), UNIFORM, and MAXIMIN for a particular reward structure. This means that for a particular reward structure in the unobserved case, the door choice made by a single subject was used against all strategies. However, notice that a choice made in a single reward structure was not used for the other three reward structures. For the 5, 20, and unlimited observation cases it was necessary to obtain 40 sample points (subject door choices) for each algorithm in each reward structure. This leads to the following statistical significance tests: Test run for the unobserved case: I ran the Brunner-Puri test for repeated observations Brunner et al. [1999] in this case. I chose to run the Brunner-Puri test because the Brunner- Puri method is better suited to non-continuous distributions with discrete data points for one-way designs. It is also necessary to use a test that deals with repeated observations since the choices made by subjects were repeated across algorithms. An alternative and more well known method to the Brunner-Puri method is the Friedman test for repeated measures, however, the Brunner-Puri method is a more robust method. Test run for the 5, 20, and unlimited observation cases: For these cases I tested each of the algorithms separately (i.e. I did not use one subject’s choice across all algorithms as in the unobserved case) so it is necessary to use a dierent type of test. Once again due to the non-normal distribution of my data I chose a more robust method for this test. In particular I chose to run Yuen’s test for comparing trimmed means Yuen [1974]. For my tests I used a standard 20% trimmed mean. A trimmed mean refers to a situation where a certain proportion of the largest and smallest sample points are removed and the remaining sample points are averaged. This is typically done to help reduce variance in data collections that may have extreme outliers that can skew data sets Wilcox [1998, 2005]. A.2 Statistical Significance Tests for MATCH In the experimental setup used in MATCH subjects were required to play all strategies for a given reward structure. This lead to a within-subjects experimental design which allowed me to directly compare the results of each strategy for each subject in each reward structure. Given that a direct comparison could be made, this allowed me to evaluate the statistical significance of my results using the bootstrap-t method [Wilcox, 2003]. The bootstrap-t method is a more robust statistical method than the classic t-test, which is necessary given the non-normal discrete distribution of my results. The bootstrap-t method also allows for direct comparison between two strategies to determine if one strategy is statistically significantly better than an alternate strategy. Thus, the use of the bootstrap-t method over the classic t-test in these experiments is justified. 140 Appendix B: Reward Structures Doors 0 1 2 3 4 5 6 7 Subject Reward 1 9 5 6 7 1 10 3 Subject Penalty -2 -4 -3 -3 -3 -2 -4 -3 Defender Reward 1 4 2 3 4 1 5 2 Defender Penalty -5 -8 -1 -6 -5 -1 -7 -7 Table B.1: Reward structure 1 Doors 0 1 2 3 4 5 6 7 Subject Reward 8 5 3 10 1 3 9 4 Subject Penalty -3 -2 -3 -2 -3 -3 -2 -3 Defender Reward 4 3 1 5 1 2 5 2 Defender Penalty -8 -10 -1 -8 -1 -3 -11 -5 Table B.2: Reward structure 2 Doors 0 1 2 3 4 5 6 7 Subject Reward 8 5 2 10 1 3 9 4 Subject Penalty -3 -3 -3 -3 -3 -3 -3 -3 Defender Reward 4 3 1 5 1 2 5 2 Defender Penalty -8 -5 -1 -10 -5 -3 -9 -6 Table B.3: Reward structure 3 141 Doors 0 1 2 3 4 5 6 7 Subject Reward 8 5 2 10 1 3 9 4 Subject Penalty -3 -3 -3 -3 -3 -3 -3 -3 Defender Reward 3 3 3 3 3 3 3 3 Defender Penalty -8 -5 -2 -10 -1 -3 -9 -4 Table B.4: Reward structure 4 Doors 0 1 2 3 4 5 6 7 Subject Reward 10 8 3 7 6 7 8 2 Subject Penalty -7 -4 -6 -8 -4 -2 -9 -3 Defender Reward 2 6 7 7 8 8 6 9 Defender Penalty -8 -10 -3 -1 -10 -5 -2 -5 Table B.5: Reward structure 5 Doors 0 1 2 3 4 5 6 7 Subject Reward 9 8 2 9 10 1 10 1 Subject Penalty -10 -1 -10 -8 -4 -10 -5 -3 Defender Reward 3 8 9 9 7 7 4 1 Defender Penalty -10 -2 -5 -1 -7 -6 -2 -1 Table B.6: Reward structure 6 Doors 0 1 2 3 4 5 6 7 Subject Reward 8 6 1 3 1 7 3 5 Subject Penalty -6 -9 -3 -7 -7 -2 -5 -2 Defender Reward 5 3 8 3 3 4 3 6 Defender Penalty -2 -5 -4 -6 -3 -10 -7 -2 Table B.7: Reward structure 7 Doors 0 1 2 3 4 5 6 7 Subject Reward 3 7 3 9 2 9 7 8 Subject Penalty -4 -8 -5 -8 -9 -4 -1 -6 Defender Reward 5 9 10 2 10 4 8 8 Defender Penalty -10 -4 -9 -3 -10 -10 -2 -5 Table B.8: Reward structure 8 Doors 0 1 2 3 4 5 6 7 Subject Reward 1 8 7 6 7 4 5 7 Subject Penalty -1 -5 -4 -6 -5 -2 -3 -6 Defender Reward 6 8 7 7 10 10 7 7 Defender Penalty -6 -7 -5 -6 -8 -10 -3 -4 Table B.9: Reward structure 9 142 Doors 0 1 2 3 4 5 6 7 Subject Reward 7 8 1 5 5 6 6 8 Subject Penalty -4 -4 -3 -1 -2 -6 -5 -6 Defender Reward 7 7 1 6 7 4 7 6 Defender Penalty -5 -6 -5 -3 -7 -1 -8 -7 Table B.10: Reward structure 10 Doors 0 1 2 3 4 5 6 7 Subject Reward 9 3 4 5 3 6 1 7 Subject Penalty -4 -1 -1 -1 -8 -2 -5 -4 Defender Reward 9 3 8 9 1 8 4 1 Defender Penalty -5 -4 -5 -1 -10 -6 -4 -9 Table B.11: Reward structure 11 Doors 0 1 2 3 4 5 6 7 Subject Reward 6 8 1 2 7 4 1 2 Subject Penalty -8 -5 -5 -2 -9 -3 -2 -1 Defender Reward 10 7 4 1 10 1 7 8 Defender Penalty -6 -5 -3 -1 -8 -5 -3 -2 Table B.12: Reward structure 12 Doors 0 1 2 3 4 5 6 7 Subject Reward 5 5 1 10 5 3 6 5 Subject Penalty -2 -1 -7 -4 -3 -3 -7 -1 Defender Reward 7 6 2 1 5 2 8 9 Defender Penalty -5 -5 -10 -6 -4 -4 -10 -9 Table B.13: Reward structure 13 Doors 0 1 2 3 4 5 6 7 Subject Reward 7 6 9 5 1 1 4 1 Subject Penalty -4 -7 -6 -7 -4 -1 -4 -2 Defender Reward 7 10 5 4 6 10 8 8 Defender Penalty -9 -5 -8 -7 -4 -10 -4 -8 Table B.14: Reward structure 14 Doors 0 1 2 3 4 5 6 7 Subject Reward 5 7 5 4 4 4 4 10 Subject Penalty -4 -6 -5 -3 -6 -4 -4 -10 Defender Reward 2 3 7 9 6 1 4 7 Defender Penalty -5 -6 -5 -6 -6 -1 -5 -7 Table B.15: Reward structure 15 143 Doors 0 1 2 3 4 5 6 7 Subject Reward 5 10 4 7 7 7 7 6 Subject Penalty -3 -10 -3 -8 -8 -4 -5 -5 Defender Reward 7 7 6 2 6 10 2 3 Defender Penalty -3 -6 -3 -2 -10 -7 -1 -4 Table B.16: Reward structure 16 Doors 0 1 2 3 4 5 6 7 Subject Reward 9 4 8 1 9 6 7 6 Subject Penalty -6 -2 -6 -1 -6 -5 -7 -3 Defender Reward 7 7 10 7 6 1 10 6 Defender Penalty -4 -3 -6 -8 -5 -1 -7 -1 Table B.17: Reward structure 17 Doors 0 1 2 3 4 5 6 7 Subject Reward 6 7 6 4 7 4 2 7 Subject Penalty -10 -4 -5 -5 -8 -5 -1 -9 Defender Reward 3 6 3 4 6 6 5 9 Defender Penalty -10 -4 -1 -6 -4 -5 -4 -9 Table B.18: Reward structure 18 Doors 0 1 2 3 4 5 6 7 Subject Reward 8 1 3 4 5 4 3 4 Subject Penalty -8 -3 -5 -5 -5 -1 -5 -8 Defender Reward 7 9 9 10 5 7 4 4 Defender Penalty -4 -3 -10 -8 -7 -4 -4 -7 Table B.19: Reward structure 19 Doors 0 1 2 3 4 5 6 7 Subject Reward 4 6 1 5 6 7 5 6 Subject Penalty -1 -7 -1 -4 -8 -9 -6 -5 Defender Reward 4 7 1 4 5 5 8 10 Defender Penalty -8 -9 -6 -1 -8 -6 -10 -6 Table B.20: Reward structure 20 Doors 0 1 2 3 4 5 6 7 Subject Reward 7 9 7 1 4 7 6 5 Subject Penalty -5 -4 -2 -2 -1 -6 -3 -3 Defender Reward 3 8 6 7 5 4 3 5 Defender Penalty -5 -10 -8 -10 -5 -1 -6 -5 Table B.21: Reward structure 21 144 Doors 0 1 2 3 4 5 6 7 Subject Reward 5 3 6 7 7 3 7 8 Subject Penalty -10 -8 -9 -4 -6 -1 -8 -6 Defender Reward 2 2 5 7 10 6 8 6 Defender Penalty -2 -7 -5 -10 -9 -5 -10 -6 Table B.22: Reward structure 22 Doors 0 1 2 3 4 5 6 7 Subject Reward 3 1 5 7 9 1 3 10 Subject Penalty -7 -2 -7 -4 -3 -8 -4 -5 Defender Reward 7 10 7 10 1 2 4 5 Defender Penalty -5 -6 -7 -3 -5 -8 -7 -5 Table B.23: Reward structure 23 Doors 0 1 2 3 4 5 6 7 Subject Reward 6 8 4 1 7 6 4 8 Subject Penalty -7 -9 -8 -5 -6 -7 -1 -3 Defender Reward 7 10 3 6 10 4 6 6 Defender Penalty -5 -5 -6 -3 -10 -7 -3 -4 Table B.24: Reward structure 24 Doors 0 1 2 3 4 5 6 7 Subject Reward 6 10 5 3 1 4 2 4 Subject Penalty -9 -8 -8 -8 -8 -7 -8 -8 Defender Reward 8 1 8 5 8 2 10 5 Defender Penalty -6 -5 -3 -4 -5 -5 -8 -1 Table B.25: Reward structure 25 Doors 0 1 2 3 4 5 6 7 Subject Reward 6 7 1 6 6 6 5 5 Subject Penalty -6 -8 -1 -6 -7 -7 -5 -5 Defender Reward 6 7 8 4 10 2 8 3 Defender Penalty -5 -5 -8 -5 -10 -4 -9 -6 Table B.26: Reward structure 26 Doors 0 1 2 3 4 5 6 7 Subject Reward 6 1 2 6 1 4 2 3 Subject Penalty -5 -4 -3 -8 -1 -7 -6 -5 Defender Reward 7 3 7 9 10 6 7 7 Defender Penalty -10 -4 -8 -9 -10 -7 -7 -10 Table B.27: Reward structure 27 145 Doors 0 1 2 3 4 5 6 7 Subject Reward 6 3 1 5 2 4 6 6 Subject Penalty -6 -5 -1 -7 -7 -6 -6 -2 Defender Reward 10 2 9 8 6 7 7 1 Defender Penalty -10 -1 -8 -9 -5 -6 -5 -1 Table B.28: Reward structure 28 Doors 0 1 2 3 4 5 6 7 Subject Reward 5 5 3 5 4 10 3 5 Subject Penalty -8 -3 -7 -3 -5 -10 -4 -7 Defender Reward 6 7 4 7 7 8 8 4 Defender Penalty -5 -6 -4 -4 -7 -6 -10 -1 Table B.29: Reward structure 29 Doors 0 1 2 3 4 5 6 7 Subject Reward 7 2 8 8 7 1 5 8 Subject Penalty -6 -4 -9 -10 -8 -1 -4 -9 Defender Reward 6 1 7 7 10 4 2 8 Defender Penalty -5 -1 -7 -9 -9 -5 -3 -8 Table B.30: Reward structure 30 Doors 0 1 2 3 4 5 6 7 Subject Reward 2 6 3 7 5 5 1 5 Subject Penalty -2 -6 -3 -6 -6 -5 -1 -7 Defender Reward 4 3 7 5 5 1 3 4 Defender Penalty -4 -5 -7 -8 -6 -1 -5 -3 Table B.31: Reward structure 31 Doors 0 1 2 3 4 5 6 7 Subject Reward 4 4 5 1 5 5 3 10 Subject Penalty -10 -4 -9 -3 -9 -4 -3 -8 Defender Reward 7 3 8 8 7 5 9 6 Defender Penalty -5 -5 -4 -1 -9 -3 -10 -5 Table B.32: Reward structure 32 Doors 0 1 2 3 4 5 6 7 Subject Reward 10 4 6 4 5 4 5 7 Subject Penalty -9 -5 -8 -5 -9 -9 -7 -10 Defender Reward 6 6 1 2 6 3 6 4 Defender Penalty -8 -5 -3 -1 -7 -7 -8 -8 Table B.33: Reward structure 33 146 Doors 0 1 2 3 4 5 6 7 Subject Reward 9 4 2 4 3 1 10 10 Subject Penalty -2 -7 -4 -4 -2 -3 -3 -10 Defender Reward 8 5 8 4 8 2 8 3 Defender Penalty -9 -10 -3 -4 -8 -4 -8 -1 Table B.34: Reward structure 34 Doors 0 1 2 3 4 5 6 7 Subject Reward 2 6 6 5 5 10 5 5 Subject Penalty -3 -7 -9 -7 -5 -10 -4 -10 Defender Reward 6 7 3 7 10 3 6 7 Defender Penalty -5 -10 -3 -4 -5 -7 -6 -5 Table B.35: Reward structure 35 Doors 0 1 2 3 4 5 6 7 Subject Reward 3 10 4 8 5 4 8 4 Subject Penalty -10 -10 -8 -6 -7 -6 -1 -1 Defender Reward 4 4 5 3 3 3 1 3 Defender Penalty -9 -7 -7 -6 -8 -10 -7 -1 Table B.36: Reward structure 36 Doors 0 1 2 3 4 5 6 7 Subject Reward 4 10 6 1 9 9 7 1 Subject Penalty -4 -4 -3 -4 -6 -1 -5 -1 Defender Reward 4 6 6 4 3 9 4 10 Defender Penalty -3 -5 -5 -10 -5 -7 -7 -2 Table B.37: Reward structure 37 Doors 0 1 2 3 4 5 6 7 Subject Reward 6 2 9 6 3 5 3 3 Subject Penalty -4 -10 -8 -1 -10 -10 -10 -8 Defender Reward 6 10 5 8 4 6 3 5 Defender Penalty -7 -4 -10 -5 -2 -6 -1 -3 Table B.38: Reward structure 38 Doors 0 1 2 3 4 5 6 7 Subject Reward 7 6 8 4 10 2 6 5 Subject Penalty -8 -10 -8 -3 -6 -2 -4 -8 Defender Reward 9 7 4 4 10 2 5 6 Defender Penalty -10 -3 -7 -3 -5 -5 -4 -3 Table B.39: Reward structure 39 147 Doors 0 1 2 3 4 5 6 7 Subject Reward 2 10 7 9 1 3 1 4 Subject Penalty -6 -2 -2 -8 -9 -7 -4 -8 Defender Reward 5 7 7 4 6 1 4 6 Defender Penalty -6 -3 -1 -7 -7 -7 -7 -7 Table B.40: Reward structure 40 Doors 0 1 2 3 4 5 6 7 Subject Reward 9 6 7 7 5 10 4 7 Subject Penalty -9 -4 -5 -7 -4 -10 -3 -4 Defender Reward 4 5 6 2 1 3 3 1 Defender Penalty -6 -6 -9 -3 -1 -5 -5 -2 Table B.41: Reward structure 41 Doors 0 1 2 3 4 5 6 7 Subject Reward 5 9 7 9 7 5 7 9 Subject Penalty -3 -8 -3 -10 -5 -1 -4 -8 Defender Reward 4 1 8 7 1 9 6 4 Defender Penalty -5 -2 -8 -8 -1 -9 -6 -7 Table B.42: Reward structure 42 Doors 0 1 2 3 4 5 6 7 Subject Reward 5 5 8 10 4 3 4 4 Subject Penalty -7 -7 -9 -10 -7 -5 -5 -6 Defender Reward 5 6 5 1 7 6 6 1 Defender Penalty -5 -7 -7 -1 -7 -8 -7 -3 Table B.43: Reward structure 43 Doors 0 1 2 3 4 5 6 7 Subject Reward 9 8 4 10 6 6 9 7 Subject Penalty -8 -7 -4 -10 -5 -5 -8 -5 Defender Reward 10 4 3 5 1 3 5 3 Defender Penalty -10 -8 -9 -8 -3 -4 -8 -8 Table B.44: Reward structure 44 Doors 0 1 2 3 4 5 6 7 Subject Reward 5 6 4 8 2 4 10 7 Subject Penalty -6 -8 -3 -5 -3 -6 -10 -5 Defender Reward 4 4 4 1 4 7 7 4 Defender Penalty -5 -5 -2 -3 -1 -6 -7 -3 Table B.45: Reward structure 45 148 Doors 0 1 2 3 4 5 6 7 Subject Reward 8 7 7 1 7 3 7 8 Subject Penalty -3 -2 -3 -1 -5 -1 -5 -6 Defender Reward 6 1 2 5 4 5 3 6 Defender Penalty -6 -1 -3 -8 -7 -6 -4 -7 Table B.46: Reward structure 46 Doors 0 1 2 3 4 5 6 7 Subject Reward 5 1 7 5 2 5 2 1 Subject Penalty -7 -4 -10 -6 -4 -7 -3 -1 Defender Reward 6 3 6 2 4 5 4 1 Defender Penalty -7 -4 -6 -1 -6 -6 -5 -2 Table B.47: Reward structure 47 Doors 0 1 2 3 4 5 6 7 Subject Reward 10 4 1 7 8 5 5 5 Subject Penalty -8 -6 -1 -8 -9 -6 -8 -4 Defender Reward 7 4 1 8 3 8 8 2 Defender Penalty -8 -5 -4 -7 -4 -7 -9 -1 Table B.48: Reward structure 48 Doors 0 1 2 3 4 5 6 7 Subject Reward 10 6 4 5 7 5 3 8 Subject Penalty -10 -7 -5 -5 -3 -7 -4 -7 Defender Reward 7 7 6 8 7 3 5 10 Defender Penalty -6 -4 -8 -9 -6 -6 -4 -10 Table B.49: Reward structure 49 Doors 0 1 2 3 4 5 6 7 Subject Reward 6 1 2 8 9 4 4 5 Subject Penalty -7 -1 -3 -10 -5 -5 -7 -4 Defender Reward 10 3 6 8 8 1 8 6 Defender Penalty -4 -2 -2 -7 -7 -1 -4 -5 Table B.50: Reward structure 50 Doors 0 1 2 3 4 5 6 7 Subject Reward 5 4 6 5 5 10 5 4 Subject Penalty -6 -6 -5 -6 -6 -10 -5 -7 Defender Reward 7 1 4 4 6 4 1 2 Defender Penalty -5 -3 -7 -4 -5 -2 -1 -3 Table B.51: Reward structure 51 149 Doors 0 1 2 3 4 5 6 7 Subject Reward 7 4 6 9 5 4 1 4 Subject Penalty -3 -2 -5 -5 -2 -1 -1 -1 Defender Reward 6 2 2 7 1 4 6 8 Defender Penalty -4 -1 -3 -7 -2 -6 -6 -5 Table B.52: Reward structure 52 Doors 0 1 2 3 4 5 6 7 Subject Reward 3 3 4 10 4 5 5 7 Subject Penalty -6 -6 -7 -10 -6 -7 -3 -10 Defender Reward 4 2 4 1 6 6 6 4 Defender Penalty -6 -1 -6 -1 -7 -7 -7 -6 Table B.53: Reward structure 53 Doors 0 1 2 3 4 5 6 7 Subject Reward 7 9 8 6 5 1 4 6 Subject Penalty -8 -8 -9 -5 -7 -1 -5 -6 Defender Reward 10 5 9 8 6 1 3 7 Defender Penalty -8 -7 -8 -4 -6 -4 -2 -7 Table B.54: Reward structure 54 Doors 0 1 2 3 4 5 6 7 Subject Reward 8 8 1 6 9 5 8 5 Subject Penalty -7 -5 -1 -6 -8 -1 -4 -2 Defender Reward 7 3 6 6 5 7 1 2 Defender Penalty -6 -9 -1 -7 -3 -2 -4 -4 Table B.55: Reward structure 55 Doors 0 1 2 3 4 5 6 7 Subject Reward 8 4 1 3 3 2 9 5 Subject Penalty -4 -3 -1 -3 -2 -1 -9 -4 Defender Reward 1 6 6 1 10 6 2 8 Defender Penalty -4 -10 -6 -3 -10 -7 -3 -7 Table B.56: Reward structure 56 Doors 0 1 2 3 4 5 6 7 Subject Reward 7 5 10 6 8 1 5 9 Subject Penalty -7 -9 -6 -2 -2 -3 -6 -6 Defender Reward 9 1 7 3 10 7 8 8 Defender Penalty -7 -5 -7 -4 -7 -5 -5 -6 Table B.57: Reward structure 57 150 Doors 0 1 2 3 4 5 6 7 Subject Reward 5 9 6 5 9 5 3 6 Subject Penalty -10 -4 -5 -1 -9 -7 -3 -5 Defender Reward 9 8 6 8 8 7 5 1 Defender Penalty -5 -6 -4 -7 -5 -6 -3 -1 Table B.58: Reward structure 58 Doors 0 1 2 3 4 5 6 7 Subject Reward 7 3 8 4 2 10 7 4 Subject Penalty -3 -4 -8 -3 -4 -10 -9 -2 Defender Reward 6 4 6 9 1 5 9 7 Defender Penalty -3 -7 -8 -8 -1 -5 -7 -7 Table B.59: Reward structure 59 Doors 0 1 2 3 4 5 6 7 Subject Reward 6 10 2 6 7 4 9 10 Subject Penalty -3 -7 -2 -4 -7 -3 -2 -10 Defender Reward 4 4 7 10 9 6 2 6 Defender Penalty -10 -5 -8 -10 -6 -8 -4 -8 Table B.60: Reward structure 60 Doors 0 1 2 3 4 5 6 7 Subject Reward 8 9 10 6 1 6 10 8 Subject Penalty -10 -3 -3 -9 -2 -4 -3 -3 Defender Reward 10 6 6 4 2 3 8 4 Defender Penalty -9 -6 -9 -2 -5 -5 -10 -6 Table B.61: Reward structure 61 Doors 0 1 2 3 4 5 6 7 Subject Reward 4 7 6 4 1 9 8 6 Subject Penalty -4 -6 -6 -8 -1 -8 -6 -6 Defender Reward 2 9 1 6 6 4 8 6 Defender Penalty -3 -7 -1 -7 -7 -5 -10 -5 Table B.62: Reward structure 62 Doors 0 1 2 3 4 5 6 7 Subject Reward 7 6 1 5 7 10 6 6 Subject Penalty -2 -7 -2 -8 -3 -4 -6 -2 Defender Reward 6 5 5 3 7 1 7 4 Defender Penalty -9 -6 -4 -4 -8 -1 -8 -4 Table B.63: Reward structure 63 151 Doors 0 1 2 3 4 5 6 7 Subject Reward 2 8 6 6 4 1 3 8 Subject Penalty -1 -10 -6 -8 -2 -2 -2 -8 Defender Reward 6 7 1 10 6 4 5 4 Defender Penalty -7 -9 -3 -9 -6 -2 -6 -4 Table B.64: Reward structure 64 Doors 0 1 2 3 4 5 6 7 Subject Reward 3 4 5 1 2 3 2 2 Subject Penalty -3 -4 -7 -3 -1 -7 -3 -4 Defender Reward 5 8 9 1 10 6 8 8 Defender Penalty -4 -5 -8 -1 -7 -2 -5 -4 Table B.65: Reward structure 65 Doors 0 1 2 3 4 5 6 7 Subject Reward 4 1 4 1 5 5 6 4 Subject Penalty -8 -1 -10 -7 -9 -6 -8 -7 Defender Reward 7 10 5 2 8 1 1 3 Defender Penalty -7 -7 -4 -4 -4 -5 -2 -7 Table B.66: Reward structure 66 Doors 0 1 2 3 4 5 6 7 Subject Reward 10 1 7 8 7 5 1 7 Subject Penalty -8 -7 -10 -6 -9 -7 -5 -6 Defender Reward 8 1 6 6 4 4 1 5 Defender Penalty -8 -1 -8 -4 -6 -3 -3 -1 Table B.67: Reward structure 67 Doors 0 1 2 3 4 5 6 7 Subject Reward 9 9 9 4 10 5 8 6 Subject Penalty -9 -6 -10 -6 -8 -3 -6 -4 Defender Reward 10 8 1 2 4 5 6 10 Defender Penalty -7 -8 -2 -3 -6 -5 -6 -8 Table B.68: Reward structure 68 Doors 0 1 2 3 4 5 6 7 Subject Reward 10 7 8 4 6 4 1 10 Subject Penalty -9 -1 -7 -7 -3 -2 -1 -4 Defender Reward 6 5 4 2 5 1 8 6 Defender Penalty -5 -5 -8 -3 -8 -7 -9 -1 Table B.69: Reward structure 69 152 Doors 0 1 2 3 4 5 6 7 Subject Reward 5 8 9 6 1 6 10 2 Subject Penalty -5 -5 -6 -10 -5 -9 -6 -5 Defender Reward 3 3 10 10 8 9 9 1 Defender Penalty -3 -4 -7 -9 -5 -5 -6 -2 Table B.70: Reward structure 70 Doors 0 1 2 3 4 5 6 7 Subject Reward 7 1 5 6 6 4 1 4 Subject Penalty -9 -1 -2 -4 -7 -3 -3 -2 Defender Reward 6 5 7 3 4 2 3 1 Defender Penalty -7 -6 -10 -5 -1 -6 -4 -2 Table B.71: Reward structure 71 Doors 0 1 2 3 4 5 6 7 Subject Reward 8 10 5 5 1 5 6 10 Subject Penalty -3 -8 -6 -7 -3 -7 -6 -6 Defender Reward 1 6 10 2 6 6 1 8 Defender Penalty -4 -6 -8 -5 -5 -4 -5 -8 Table B.72: Reward structure 72 Doors 0 1 2 3 4 5 6 7 Subject Reward 4 8 3 8 7 7 10 8 Subject Penalty -4 -6 -3 -6 -7 -4 -10 -9 Defender Reward 3 7 4 8 9 3 1 10 Defender Penalty -5 -6 -5 -7 -7 -6 -6 -7 Table B.73: Reward structure 73 Doors 0 1 2 3 4 5 6 7 Subject Reward 8 1 2 5 5 2 3 8 Subject Penalty -2 -2 -1 -4 -6 -3 -7 -4 Defender Reward 7 8 2 10 9 6 5 1 Defender Penalty -6 -2 -5 -4 -7 -3 -6 -1 Table B.74: Reward structure 74 Doors 0 1 2 3 4 5 6 7 Subject Reward 2 8 8 6 3 6 9 1 Subject Penalty -5 -7 -9 -7 -5 -2 -7 -1 Defender Reward 10 8 8 6 9 4 6 4 Defender Penalty -10 -3 -8 -7 -7 -7 -1 -7 Table B.75: Reward structure 75 153 Doors 0 1 2 3 4 5 6 7 Subject Reward 7 9 3 7 10 5 5 1 Subject Penalty -8 -5 -4 -9 -7 -5 -7 -1 Defender Reward 8 5 5 10 9 8 9 3 Defender Penalty -5 -7 -6 -10 -6 -9 -6 -7 Table B.76: Reward structure 76 Doors 0 1 2 3 4 5 6 7 Subject Reward 6 6 7 8 1 9 9 7 Subject Penalty -6 -9 -2 -8 -1 -8 -10 -5 Defender Reward 6 3 10 7 9 1 6 3 Defender Penalty -8 -8 -5 -6 -7 -4 -3 -7 Table B.77: Reward structure 77 Doors 0 1 2 3 4 5 6 7 Subject Reward 8 6 1 10 9 5 4 8 Subject Penalty -3 -5 -2 -9 -5 -2 -5 -7 Defender Reward 1 2 6 9 6 8 1 6 Defender Penalty -3 -1 -6 -8 -2 -5 -5 -5 Table B.78: Reward structure 78 Doors 0 1 2 3 4 5 6 7 Subject Reward 6 7 1 5 5 6 4 5 Subject Penalty -4 -3 -5 -8 -2 -1 -7 -5 Defender Reward 4 8 10 5 5 8 7 7 Defender Penalty -3 -6 -5 -3 -3 -6 -10 -4 Table B.79: Reward structure 79 Doors 0 1 2 3 4 5 6 7 Subject Reward 1 5 5 10 4 5 6 3 Subject Penalty -6 -5 -4 -6 -6 -1 -6 -5 Defender Reward 1 6 7 5 10 5 4 6 Defender Penalty -6 -5 -8 -3 -4 -4 -2 -5 Table B.80: Reward structure 80 Doors 0 1 2 3 4 5 6 7 Subject Reward 2 1 7 7 1 6 3 9 Subject Penalty -5 -1 -2 -6 -9 -9 -9 -10 Defender Reward 2 10 9 4 6 9 4 10 Defender Penalty -1 -8 -9 -5 -10 -10 -3 -2 Table B.81: Reward structure 81 154 Doors 0 1 2 3 4 5 6 7 Subject Reward 5 6 5 7 8 6 5 5 Subject Penalty -7 -5 -1 -3 -10 -7 -4 -8 Defender Reward 3 1 6 10 4 2 3 4 Defender Penalty -5 -8 -5 -8 -3 -6 -4 -10 Table B.82: Reward structure 82 Doors 0 1 2 3 4 5 6 7 Subject Reward 7 6 1 8 6 4 9 9 Subject Penalty -4 -2 -1 -2 -2 -2 -5 -7 Defender Reward 1 2 2 3 8 4 5 3 Defender Penalty -5 -2 -7 -1 -8 -9 -10 -8 Table B.83: Reward structure 83 Doors 0 1 2 3 4 5 6 7 Subject Reward 7 8 6 4 6 9 1 2 Subject Penalty -7 -5 -6 -10 -3 -6 -5 -1 Defender Reward 6 6 6 5 5 3 5 1 Defender Penalty -7 -4 -5 -7 -1 -9 -6 -5 Table B.84: Reward structure 84 Doors 0 1 2 3 4 5 6 7 Subject Reward 5 1 6 7 2 6 4 6 Subject Penalty -8 -1 -5 -2 -1 -2 -2 -5 Defender Reward 4 5 4 8 1 10 4 2 Defender Penalty -8 -7 -5 -10 -7 -8 -7 -6 Table B.85: Reward structure 85 Doors 0 1 2 3 4 5 6 7 Subject Reward 9 10 9 1 2 6 9 7 Subject Penalty -5 -1 -2 -5 -6 -3 -7 -9 Defender Reward 5 9 4 2 7 8 5 4 Defender Penalty -3 -6 -1 -10 -3 -4 -2 -2 Table B.86: Reward structure 86 Doors 0 1 2 3 4 5 6 7 Subject Reward 1 6 4 4 7 6 7 6 Subject Penalty -2 -5 -7 -1 -5 -2 -1 -3 Defender Reward 1 3 5 2 5 4 6 5 Defender Penalty -1 -5 -8 -2 -7 -8 -4 -5 Table B.87: Reward structure 87 155 Doors 0 1 2 3 4 5 6 7 Subject Reward 3 10 3 5 2 4 9 1 Subject Penalty -8 -7 -5 -3 -6 -10 -7 -2 Defender Reward 6 6 6 6 4 8 10 1 Defender Penalty -5 -7 -5 -10 -4 -7 -7 -8 Table B.88: Reward structure 88 Doors 0 1 2 3 4 5 6 7 Subject Reward 1 7 8 8 9 8 4 7 Subject Penalty -3 -6 -2 -7 -6 -1 -5 -4 Defender Reward 5 3 6 3 6 6 1 8 Defender Penalty -7 -9 -5 -10 -4 -7 -8 -1 Table B.89: Reward structure 89 Doors 0 1 2 3 4 5 6 7 Subject Reward 9 3 1 5 4 6 1 7 Subject Penalty -9 -7 -1 -5 -5 -9 -3 -4 Defender Reward 9 5 3 10 8 4 3 3 Defender Penalty -8 -10 -9 -9 -9 -1 -5 -9 Table B.90: Reward structure 90 Doors 0 1 2 3 4 5 6 7 Subject Reward 8 9 6 6 5 5 1 7 Subject Penalty -8 -3 -5 -5 -1 -3 -2 -4 Defender Reward 6 1 4 6 7 5 6 1 Defender Penalty -3 -1 -6 -1 -8 -4 -2 -2 Table B.91: Reward structure 91 Doors 0 1 2 3 4 5 6 7 Subject Reward 8 1 9 7 8 9 10 9 Subject Penalty -5 -2 -5 -2 -6 -5 -8 -9 Defender Reward 8 8 7 6 6 10 8 8 Defender Penalty -7 -7 -2 -9 -10 -7 -3 -5 Table B.92: Reward structure 92 Doors 0 1 2 3 4 5 6 7 Subject Reward 4 7 5 9 6 1 5 4 Subject Penalty -4 -3 -4 -1 -3 -5 -2 -7 Defender Reward 7 10 2 10 6 3 6 7 Defender Penalty -4 -6 -4 -6 -10 -5 -5 -6 Table B.93: Reward structure 93 156 Doors 0 1 2 3 4 5 6 7 Subject Reward 10 3 6 5 1 8 1 6 Subject Penalty -8 -5 -4 -8 -8 -3 -1 -2 Defender Reward 3 5 3 1 6 4 8 4 Defender Penalty -9 -7 -1 -4 -2 -9 -6 -2 Table B.94: Reward structure 94 Doors 0 1 2 3 4 5 6 7 Subject Reward 6 1 8 9 10 8 6 3 Subject Penalty -4 -7 -2 -8 -1 -6 -10 -4 Defender Reward 5 5 4 1 10 1 2 6 Defender Penalty -8 -8 -5 -7 -8 -5 -7 -6 Table B.95: Reward structure 95 Doors 0 1 2 3 4 5 6 7 Subject Reward 6 7 8 7 6 8 5 6 Subject Penalty -1 -10 -5 -9 -5 -6 -9 -5 Defender Reward 10 5 5 2 6 3 6 4 Defender Penalty -7 -5 -2 -10 -9 -5 -10 -5 Table B.96: Reward structure 96 Doors 0 1 2 3 4 5 6 7 Subject Reward 8 4 2 3 5 10 5 4 Subject Penalty -6 -8 -1 -2 -6 -6 -3 -10 Defender Reward 5 3 4 2 5 4 1 2 Defender Penalty -7 -4 -7 -5 -1 -4 -6 -6 Table B.97: Reward structure 97 Doors 0 1 2 3 4 5 6 7 Subject Reward 2 4 4 1 9 6 8 5 Subject Penalty -6 -2 -3 -1 -4 -6 -10 -6 Defender Reward 10 8 6 2 6 10 7 9 Defender Penalty -4 -6 -9 -7 -1 -10 -4 -5 Table B.98: Reward structure 98 Doors 0 1 2 3 4 5 6 7 Subject Reward 7 4 4 4 8 8 1 10 Subject Penalty -8 -4 -8 -1 -1 -6 -5 -6 Defender Reward 10 5 4 9 5 3 4 2 Defender Penalty -10 -1 -5 -7 -5 -1 -4 -2 Table B.99: Reward structure 99 157 Doors 0 1 2 3 4 5 6 7 Subject Reward 6 6 1 7 2 7 4 6 Subject Penalty -8 -3 -7 -9 -6 -5 -1 -6 Defender Reward 1 4 8 5 10 7 2 8 Defender Penalty -6 -7 -8 -2 -7 -2 -9 -3 Table B.100: Reward structure 100 Doors 0 1 2 3 4 5 6 7 Subject Reward 6 1 4 10 1 1 2 6 Subject Penalty -7 -10 -7 -1 -6 -4 -4 -7 Defender Reward 6 8 1 8 10 1 4 3 Defender Penalty -6 -7 -4 -1 -7 -5 -7 -5 Table B.101: Reward structure 101 Doors 0 1 2 3 4 5 6 7 Subject Reward 2 1 9 10 7 1 6 6 Subject Penalty -6 -8 -2 -4 -7 -7 -4 -3 Defender Reward 10 5 6 5 7 4 8 1 Defender Penalty -3 -5 -8 -5 -5 -1 -4 -3 Table B.102: Reward structure 102 Doors 0 1 2 3 4 5 6 7 Subject Reward 7 10 2 9 9 1 1 8 Subject Penalty -2 -3 -4 -4 -2 -5 -6 -3 Defender Reward 7 10 4 5 8 1 6 7 Defender Penalty -2 -7 -6 -5 -10 -2 -7 -5 Table B.103: Reward structure 103 Doors 0 1 2 3 4 5 6 7 Subject Reward 4 6 5 8 10 4 10 8 Subject Penalty -3 -5 -10 -4 -5 -3 -7 -4 Defender Reward 3 4 4 4 5 10 6 6 Defender Penalty -5 -6 -6 -6 -6 -10 -10 -8 Table B.104: Reward structure 104 Doors 0 1 2 3 4 5 6 7 Subject Reward 4 8 4 10 8 1 6 5 Subject Penalty -3 -6 -3 -6 -4 -7 -1 -5 Defender Reward 3 8 5 10 6 1 5 5 Defender Penalty -1 -3 -3 -6 -4 -7 -7 -2 Table B.105: Reward structure 105 158 Doors 0 1 2 3 4 5 6 7 Subject Reward 2 4 7 9 9 8 1 10 Subject Penalty -5 -2 -4 -6 -2 -4 -3 -3 Defender Reward 4 1 3 8 4 10 7 10 Defender Penalty -9 -5 -3 -3 -9 -8 -5 -9 Table B.106: Reward structure 106 Doors 0 1 2 3 4 5 6 7 Subject Reward 3 10 5 8 7 8 6 1 Subject Penalty -5 -3 -6 -4 -1 -2 -4 -7 Defender Reward 3 9 6 7 10 6 7 1 Defender Penalty -7 -4 -4 -7 -8 -5 -1 -10 Table B.107: Reward structure 107 Doors 0 1 2 3 4 5 6 7 Subject Reward 7 9 3 2 4 7 5 5 Subject Penalty -10 -1 -7 -3 -4 -8 -2 -4 Defender Reward 10 10 5 9 9 8 1 7 Defender Penalty -7 -3 -3 -8 -7 -3 -8 -5 Table B.108: Reward structure 108 159 Appendix C: Strategies DOBSS COBRA COBRA COBRA COBRA MAXIMIN (0,2.5) (1,2.5) (.37,2.5) (.03,2.5) Door 1 0 .06 0 0 .06 .56 Door 2 .58 .63 .77 .64 .63 .53 Door 3 .45 .21 0 .23 .21 0 Door 4 .51 .58 .81 .63 .58 .48 Door 5 .56 .51 .70 .52 .51 .37 Door 6 0 .06 0 0 .06 0 Door 7 .61 .55 .69 .55 .55 .44 Door 8 .27 .36 0 .40 .36 .59 Table C.1: Mixed strategies for reward structure 1 DOBSS COBRA COBRA COBRA COBRA MAXIMIN (0,2.5) (1,2.5) (.54,2.5) (.41,2.5) Door 1 .55 .57 1 .67 .61 .53 Door 2 .44 .55 0 .60 .56 .64 Door 3 .18 0 0 .05 .13 0 Door 4 .67 .53 1 .62 .57 .48 Door 5 0 0 0 0 0 0 Door 6 .18 .31 0 .05 .13 .27 Door 7 .64 .61 1 .69 .65 .58 Door 8 .30 .41 0 .29 .32 .48 Table C.2: Mixed strategies for reward structure 2 160 DOBSS COBRA COBRA COBRA COBRA MAXIMIN (0,2.5) (1,2.5) (.75,2.5) (.25,2.5) Door 1 .58 .52 1 .71 .55 .48 Door 2 .43 .41 0 .60 .46 .35 Door 3 .09 0 0 0 0 0 Door 4 .65 .55 1 .70 .57 .52 Door 5 0 .17 0 0 0 .47 Door 6 .24 .26 0 0 .33 .17 Door 7 .62 .52 1 .68 .54 .49 Door 8 .35 .53 0 .28 .51 .48 Table C.3: Mixed strategies for reward structure 3 DOBSS COBRA COBRA COBRA COBRA MAXIMIN (0,2.5) (1,2.5) (.75,2.5) (.25,2.5) Door 1 .58 .58 1 .68 .58 .58 Door 2 .43 .43 0 .56 .43 .43 Door 3 .09 .09 0 0 .09 .09 Door 4 .65 .65 1 .73 .65 .65 Door 5 0 0 0 0 0 0 Door 6 .24 .24 0 0 .24 .24 Door 7 .62 .62 1 .70 .62 .62 Door 8 .35 .35 0 .32 .35 .35 Table C.4: Mixed strategies for reward structure 4 DOBSS BRQR MATCH Door 1 0.49118 0.56923 0.57388 Door 2 0.52917 0.57955 0.55339 Door 3 0.15 0.18303 0.18394 Door 4 0.35667 0.20853 0.2389 Door 5 0.435 0.5053 0.48196 Door 6 0.59445 0.47195 0.43158 Door 7 0.37353 0.29801 0.29979 Door 8 0.070004 0.18439 0.23657 Table C.5: Mixed strategies for reward structure 5 161 DOBSS BRQR MATCH Door 1 0.42135 0.53544 0.55473 Door 2 0.7784 0.51819 0.46059 Door 3 0.083799 0.20815 0.2212 Door 4 0.47092 0.35989 0.32412 Door 5 0.64326 0.63863 0.56254 Door 6 0.00050859 0.15929 0.23964 Door 7 0.60037 0.58041 0.51197 Door 8 0.0013986 0 0.12521 Table C.6: Mixed strategies for reward structure 6 DOBSS BRQR MATCH Door 1 0.53446 0.36491 0.39159 Door 2 0.36549 0.42707 0.40101 Door 3 0.1206 0.19989 0.20146 Door 4 0.24824 0.35737 0.38017 Door 5 0.0603 0.12692 0.15881 Door 6 0.72027 0.72366 0.66188 Door 7 0.3103 0.42603 0.45685 Door 8 0.64034 0.37415 0.34822 Table C.7: Mixed strategies for reward structure 7 DOBSS BRQR MATCH Door 1 0.21734 0.34625 0.44487 Door 2 0.36809 0.33067 0.27811 Door 3 0.19017 0.29879 0.32545 Door 4 0.44244 0.43861 0.39942 Door 5 0.047399 0.19562 0.28346 Door 6 0.57857 0.61677 0.58471 Door 7 0.69017 0.35675 0.32151 Door 8 0.46581 0.41654 0.36249 Table C.8: Mixed strategies for reward structure 8 162 BRQR MATCH Door 1 0.031979 0.30562 Door 2 0.4777 0.43852 Door 3 0.43569 0.40342 Door 4 0.40438 0.37115 Door 5 0.45842 0.40929 Door 6 0.49791 0.43379 Door 7 0.32578 0.29326 Door 8 0.36814 0.34495 Table C.9: Mixed strategies for reward structure 9 BRQR MATCH Door 1 0.44102 0.3917 Door 2 0.48892 0.44037 Door 3 0.0083501 0.30092 Door 4 0.39495 0.33395 Door 5 0.49699 0.42901 Door 6 0.21231 0.23583 Door 7 0.46688 0.42343 Door 8 0.49058 0.44478 Table C.10: Mixed strategies for reward structure 10 BRQR MATCH Door 1 0.46554 0.41412 Door 2 0.33212 0.3801 Door 3 0.44026 0.3434 Door 4 0.29086 0.19882 Door 5 0.31655 0.46278 Door 6 0.50599 0.41732 Door 7 0.038925 0.15579 Door 8 0.60976 0.62767 Table C.11: Mixed strategies for reward structure 11 163 BRQR MATCH Door 1 0.4173 0.37247 Door 2 0.50216 0.48697 Door 3 0.20484 0.24417 Door 4 0.16726 0.36237 Door 5 0.46177 0.41689 Door 6 0.59813 0.62879 Door 7 0.27382 0.24417 Door 8 0.37471 0.24417 Table C.12: Mixed strategies for reward structure 12 BRQR MATCH Door 1 0.37698 0.32346 Door 2 0.4177 0.36151 Door 3 0.14489 0.35729 Door 4 0.59455 0.57837 Door 5 0.32782 0.30269 Door 6 0.20829 0.26215 Door 7 0.4157 0.3918 Door 8 0.51406 0.42274 Table C.13: Mixed strategies for reward structure 13 BRQR MATCH Door 1 0.54347 0.50963 Door 2 0.32934 0.31286 Door 3 0.54583 0.52715 Door 4 0.42059 0.42435 Door 5 0.11906 0.18401 Door 6 0.42951 0.39819 Door 7 0.31543 0.28801 Door 8 0.29676 0.3558 Table C.14: Mixed strategies for reward structure 14 164 BRQR MATCH Door 1 0.46362 0.46477 Door 2 0.49376 0.47437 Door 3 0.37271 0.33801 Door 4 0.40476 0.33801 Door 5 0.36136 0.33801 Door 6 0.062786 0.24363 Door 7 0.39218 0.3786 Door 8 0.44883 0.4246 Table C.15: Mixed strategies for reward structure 15 BRQR MATCH Door 1 0.32818 0.30559 Door 2 0.4332 0.40911 Door 3 0.2943 0.28128 Door 4 0.33182 0.34213 Door 5 0.46742 0.46776 Door 6 0.45422 0.41073 Door 7 0.27545 0.3667 Door 8 0.4154 0.4167 Table C.16: Mixed strategies for reward structure 16 BRQR MATCH Door 1 0.44183 0.4142 Door 2 0.35812 0.29807 Door 3 0.44112 0.39231 Door 4 0.18538 0.39819 Door 5 0.49288 0.45266 Door 6 0.35812 0.36686 Door 7 0.43102 0.37965 Door 8 0.29154 0.29807 Table C.17: Mixed strategies for reward structure 17 165 BRQR MATCH Door 1 0.4458 0.47245 Door 2 0.43025 0.41434 Door 3 0.22804 0.3134 Door 4 0.41189 0.40532 Door 5 0.37431 0.34804 Door 6 0.36769 0.33505 Door 7 0.30076 0.30842 Door 8 0.44124 0.40297 Table C.18: Mixed strategies for reward structure 18 BRQR MATCH Door 1 0.37952 0.38886 Door 2 0.13566 0.15621 Door 3 0.4243 0.4259 Door 4 0.42282 0.38886 Door 5 0.48221 0.47724 Door 6 0.45763 0.40621 Door 7 0.31536 0.34371 Door 8 0.3825 0.41301 Table C.19: Mixed strategies for reward structure 19 BRQR MATCH Door 1 0.60075 0.51171 Door 2 0.45821 0.40342 Door 3 0.096945 0.411 Door 4 0.14126 0.19279 Door 5 0.44011 0.39626 Door 6 0.41305 0.35922 Door 7 0.45771 0.40342 Door 8 0.39197 0.32219 Table C.20: Mixed strategies for reward structure 20 166 BRQR MATCH Door 1 0.43642 0.39444 Door 2 0.54295 0.48029 Door 3 0.54423 0.47343 Door 4 0.10386 0.34444 Door 5 0.38067 0.32592 Door 6 0.13658 0.21605 Door 7 0.47641 0.43827 Door 8 0.37888 0.32716 Table C.21: Mixed strategies for reward structure 21 BRQR MATCH Door 1 0.16548 0.21549 Door 2 0.28342 0.35472 Door 3 0.3309 0.32378 Door 4 0.53527 0.50337 Door 5 0.44191 0.4092 Door 6 0.36485 0.33963 Door 7 0.4468 0.4271 Door 8 0.43137 0.42671 Table C.22: Mixed strategies for reward structure 22 BRQR MATCH Door 1 0.27491 0.24258 Door 2 0.19842 0.22825 Door 3 0.40263 0.3591 Door 4 0.33373 0.30569 Door 5 0.67307 0.62982 Door 6 0.18927 0.33351 Door 7 0.39357 0.40759 Door 8 0.5344 0.49347 Table C.23: Mixed strategies for reward structure 23 167 BRQR MATCH Door 1 0.38948 0.3638 Door 2 0.36987 0.34672 Door 3 0.34544 0.38547 Door 4 0.064649 0.13966 Door 5 0.49677 0.45742 Door 6 0.4578 0.46229 Door 7 0.39784 0.36393 Door 8 0.47815 0.48071 Table C.24: Mixed strategies for reward structure 24 BRQR MATCH Door 1 0.43008 0.40137 Door 2 0.62192 0.60998 Door 3 0.31931 0.31832 Door 4 0.34255 0.33198 Door 5 0.25743 0.25635 Door 6 0.46621 0.47998 Door 7 0.359 0.34427 Door 8 0.2035 0.25776 Table C.25: Mixed strategies for reward structure 25 BRQR MATCH Door 1 0.35358 0.3495 Door 2 0.33268 0.33476 Door 3 0.289 0.33548 Door 4 0.38236 0.38279 Door 5 0.4286 0.39511 Door 6 0.35396 0.37045 Door 7 0.44104 0.40884 Door 8 0.41879 0.42308 Table C.26: Mixed strategies for reward structure 26 168 BRQR MATCH Door 1 0.50705 0.49056 Door 2 0.11007 0.22798 Door 3 0.38917 0.38679 Door 4 0.41173 0.39799 Door 5 0.51401 0.39708 Door 6 0.35927 0.36399 Door 7 0.28498 0.30617 Door 8 0.42372 0.42943 Table C.27: Mixed strategies for reward structure 27 BRQR MATCH Door 1 0.49902 0.44633 Door 2 0.13916 0.20752 Door 3 0.47985 0.3833 Door 4 0.45827 0.42354 Door 5 0.26543 0.26414 Door 6 0.39678 0.36012 Door 7 0.41923 0.38678 Door 8 0.34225 0.52827 Table C.28: Mixed strategies for reward structure 28 BRQR MATCH Door 1 0.3575 0.34615 Door 2 0.47803 0.44322 Door 3 0.27061 0.29486 Door 4 0.40718 0.38461 Door 5 0.41499 0.40468 Door 6 0.43451 0.42081 Door 7 0.44495 0.4523 Door 8 0.19222 0.25338 Table C.29: Mixed strategies for reward structure 29 169 BRQR MATCH Door 1 0.43063 0.41385 Door 2 0 0.11656 Door 3 0.45197 0.41718 Door 4 0.47723 0.43919 Door 5 0.45622 0.40978 Door 6 0.31171 0.3575 Door 7 0.41129 0.42375 Door 8 0.46095 0.4222 Table C.30: Mixed strategies for reward structure 30 BRQR MATCH Door 1 0.34639 0.30989 Door 2 0.4706 0.43594 Door 3 0.46481 0.38594 Door 4 0.53714 0.48918 Door 5 0.44229 0.39631 Door 6 0.12073 0.30989 Door 7 0.30095 0.37187 Door 8 0.3171 0.30099 Table C.31: Mixed strategies for reward structure 31 BRQR MATCH Door 1 0.31577 0.30768 Door 2 0.47374 0.49997 Door 3 0.32059 0.30768 Door 4 0.069525 0.07689 Door 5 0.43191 0.43332 Door 6 0.39255 0.41174 Door 7 0.52478 0.47998 Door 8 0.47112 0.48274 Table C.32: Mixed strategies for reward structure 32 170 BRQR MATCH Door 1 0.50682 0.47617 Door 2 0.37494 0.33568 Door 3 0.37317 0.37297 Door 4 0.068584 0.22612 Door 5 0.39025 0.35976 Door 6 0.37813 0.37885 Door 7 0.44578 0.41206 Door 8 0.46232 0.4384 Table C.33: Mixed strategies for reward structure 33 BRQR MATCH Door 1 0.61702 0.55101 Door 2 0.41251 0.43955 Door 3 0.18253 0.14284 Door 4 0.37428 0.33926 Door 5 0.46092 0.40134 Door 6 0.050871 0.24282 Door 7 0.58041 0.53201 Door 8 0.32146 0.35118 Table C.34: Mixed strategies for reward structure 34 BRQR MATCH Door 1 0.28671 0.31944 Door 2 0.47916 0.47037 Door 3 0.31598 0.33862 Door 4 0.30974 0.30917 Door 5 0.34622 0.32444 Door 6 0.50902 0.5037 Door 7 0.44492 0.43386 Door 8 0.30825 0.30041 Table C.35: Mixed strategies for reward structure 35 171 BRQR MATCH Door 1 0.27227 0.30417 Door 2 0.45538 0.4164 Door 3 0.30964 0.28786 Door 4 0.46744 0.43081 Door 5 0.40012 0.38733 Door 6 0.40898 0.43081 Door 7 0.67703 0.64168 Door 8 0.0091507 0.10095 Table C.36: Mixed strategies for reward structure 36 BRQR MATCH Door 1 0.28789 0.26184 Door 2 0.51608 0.4771 Door 3 0.45695 0.39638 Door 4 0.15973 0.41724 Door 5 0.51787 0.47511 Door 6 0.55388 0.49722 Door 7 0.5076 0.47511 Door 8 0 0 Table C.37: Mixed strategies for reward structure 37 BRQR MATCH Door 1 0.56642 0.54596 Door 2 0.21141 0.21373 Door 3 0.58391 0.57991 Door 4 0.56425 0.52785 Door 5 0.22608 0.23984 Door 6 0.38485 0.391 Door 7 0.1877 0.20924 Door 8 0.27538 0.29247 Table C.38: Mixed strategies for reward structure 38 172 BRQR MATCH Door 1 0.46496 0.43254 Door 2 0.29043 0.25794 Door 3 0.49292 0.47061 Door 4 0.34785 0.33618 Door 5 0.40875 0.40989 Door 6 0.26522 0.42786 Door 7 0.43867 0.4056 Door 8 0.2912 0.25938 Table C.39: Mixed strategies for reward structure 39 BRQR MATCH Door 1 0.28949 0.30123 Door 2 0.5489 0.48743 Door 3 0.42605 0.33667 Door 4 0.55799 0.49012 Door 5 0.1915 0.24884 Door 6 0.37062 0.42907 Door 7 0.23361 0.35771 Door 8 0.38184 0.34893 Table C.40: Mixed strategies for reward structure 40 BRQR MATCH Door 1 0.45055 0.40924 Door 2 0.4554 0.4028 Door 3 0.51544 0.46143 Door 4 0.35551 0.33993 Door 5 0.04993 0.22352 Door 6 0.44562 0.40924 Door 7 0.39275 0.36392 Door 8 0.3348 0.38991 Table C.41: Mixed strategies for reward structure 41 173 BRQR MATCH Door 1 0.35292 0.34806 Door 2 0.31637 0.34585 Door 3 0.46813 0.41988 Door 4 0.40765 0.37991 Door 5 0.1051 0.27979 Door 6 0.48265 0.41321 Door 7 0.42004 0.3877 Door 8 0.44713 0.42561 Table C.42: Mixed strategies for reward structure 42 BRQR MATCH Door 1 0.37769 0.34431 Door 2 0.41657 0.38299 Door 3 0.46925 0.43362 Door 4 0.21833 0.38977 Door 5 0.37363 0.34299 Door 6 0.39878 0.38977 Door 7 0.42408 0.38977 Door 8 0.32168 0.32678 Table C.43: Mixed strategies for reward structure 43 BRQR MATCH Door 1 0.41457 0.38161 Door 2 0.43524 0.41184 Door 3 0.38047 0.40598 Door 4 0.42115 0.39756 Door 5 0.19463 0.27464 Door 6 0.26696 0.28442 Door 7 0.42835 0.40398 Door 8 0.45862 0.43998 Table C.44: Mixed strategies for reward structure 44 174 BRQR MATCH Door 1 0.42465 0.40677 Door 2 0.418 0.39719 Door 3 0.29717 0.3181 Door 4 0.52491 0.53737 Door 5 0.07314 0.11353 Door 6 0.38538 0.35371 Door 7 0.47161 0.44516 Door 8 0.40513 0.42817 Table C.45: Mixed strategies for reward structure 45 BRQR MATCH Door 1 0.49474 0.42774 Door 2 0.18393 0.34892 Door 3 0.43068 0.38921 Door 4 0.10799 0.32254 Door 5 0.4949 0.42774 Door 6 0.39571 0.32254 Door 7 0.42113 0.3599 Door 8 0.47092 0.40141 Table C.46: Mixed strategies for reward structure 46 BRQR MATCH Door 1 0.48397 0.42384 Door 2 0.31742 0.29966 Door 3 0.44684 0.39986 Door 4 0.18649 0.32828 Door 5 0.46564 0.41225 Door 6 0.47309 0.41722 Door 7 0.47019 0.39971 Door 8 0.15635 0.31919 Table C.47: Mixed strategies for reward structure 47 175 BRQR MATCH Door 1 0.51531 0.47046 Door 2 0.39659 0.34343 Door 3 0.16129 0.36075 Door 4 0.44093 0.38417 Door 5 0.43653 0.39688 Door 6 0.43674 0.36635 Door 7 0.43931 0.38417 Door 8 0.1733 0.29377 Table C.48: Mixed strategies for reward structure 48 BRQR MATCH Door 1 0.40117 0.39587 Door 2 0.29662 0.29433 Door 3 0.37767 0.39408 Door 4 0.42459 0.40977 Door 5 0.45114 0.43756 Door 6 0.36206 0.38399 Door 7 0.22716 0.25399 Door 8 0.45958 0.4304 Table C.49: Mixed strategies for reward structure 49 BRQR MATCH Door 1 0.3721 0.33997 Door 2 0.12554 0.31133 Door 3 0.27868 0.24456 Door 4 0.47045 0.42967 Door 5 0.55222 0.52342 Door 6 0.3385 0.37994 Door 7 0.35658 0.31214 Door 8 0.50593 0.45896 Table C.50: Mixed strategies for reward structure 50 176 BRQR MATCH Door 1 0.39568 0.35102 Door 2 0.35783 0.36239 Door 3 0.5387 0.50334 Door 4 0.40083 0.37229 Door 5 0.40797 0.36698 Door 6 0.32015 0.38744 Door 7 0.25174 0.33946 Door 8 0.3271 0.31709 Table C.51: Mixed strategies for reward structure 51 BRQR MATCH Door 1 0.43382 0.40477 Door 2 0.16767 0.23283 Door 3 0.40267 0.38097 Door 4 0.51126 0.46769 Door 5 0.37703 0.40954 Door 6 0.5263 0.47303 Door 7 0.13348 0.29253 Door 8 0.44777 0.33864 Table C.52: Mixed strategies for reward structure 52 BRQR MATCH Door 1 0.37649 0.36842 Door 2 0.11258 0.16666 Door 3 0.40139 0.38095 Door 4 0.26258 0.40909 Door 5 0.4276 0.3913 Door 6 0.43927 0.4 Door 7 0.53929 0.47619 Door 8 0.44081 0.4074 Table C.53: Mixed strategies for reward structure 53 177 BRQR MATCH Door 1 0.43673 0.38351 Door 2 0.51469 0.4709 Door 3 0.44994 0.40165 Door 4 0.36181 0.33287 Door 5 0.4091 0.36067 Door 6 0.096859 0.37942 Door 7 0.26558 0.26114 Door 8 0.46528 0.40985 Table C.54: Mixed strategies for reward structure 54 BRQR MATCH Door 1 0.42172 0.3958 Door 2 0.56386 0.5633 Door 3 0 0 Door 4 0.4175 0.4033 Door 5 0.35006 0.3633 Door 6 0.29089 0.27216 Door 7 0.52383 0.53426 Door 8 0.43214 0.46788 Table C.55: Mixed strategies for reward structure 55 BRQR MATCH Door 1 0.48509 0.51224 Door 2 0.50724 0.46557 Door 3 0.18479 0.26486 Door 4 0.20747 0.27081 Door 5 0.47907 0.38832 Door 6 0.40002 0.35675 Door 7 0.32164 0.37861 Door 8 0.41469 0.36284 Table C.56: Mixed strategies for reward structure 56 178 BRQR MATCH Door 1 0.40729 0.36784 Door 2 0.32416 0.35175 Door 3 0.50221 0.46784 Door 4 0.47732 0.46901 Door 5 0.49512 0.44574 Door 6 0.019537 0.18969 Door 7 0.32669 0.29313 Door 8 0.44767 0.415 Table C.57: Mixed strategies for reward structure 57 BRQR MATCH Door 1 0.28892 0.27302 Door 2 0.48979 0.47843 Door 3 0.38726 0.37703 Door 4 0.53269 0.47227 Door 5 0.39921 0.38444 Door 6 0.36113 0.3567 Door 7 0.24213 0.27983 Door 8 0.29886 0.37828 Table C.58: Mixed strategies for reward structure 58 BRQR MATCH Door 1 0.38324 0.4046 Door 2 0.37355 0.42708 Door 3 0.47918 0.45625 Door 4 0.44904 0.40364 Door 5 0 0.085923 Door 6 0.44161 0.42291 Door 7 0.39608 0.36523 Door 8 0.4773 0.43437 Table C.59: Mixed strategies for reward structure 59 179 BRQR MATCH Door 1 0.50092 0.48407 Door 2 0.38745 0.38975 Door 3 0.17799 0.27019 Door 4 0.42356 0.37112 Door 5 0.31399 0.28047 Door 6 0.34873 0.33969 Door 7 0.43053 0.47844 Door 8 0.41684 0.38628 Table C.60: Mixed strategies for reward structure 60 BRQR MATCH Door 1 0.3745 0.34318 Door 2 0.4709 0.44574 Door 3 0.56405 0.52492 Door 4 0.17164 0.17609 Door 5 0 0.16978 Door 6 0.3798 0.3721 Door 7 0.55047 0.50638 Door 8 0.48864 0.4618 Table C.61: Mixed strategies for reward structure 61 BRQR MATCH Door 1 0.33695 0.33335 Door 2 0.43828 0.39081 Door 3 0.15886 0.30954 Door 4 0.3579 0.33334 Door 5 0.30991 0.35557 Door 6 0.464 0.4359 Door 7 0.53148 0.47917 Door 8 0.40261 0.36233 Table C.62: Mixed strategies for reward structure 62 180 BRQR MATCH Door 1 0.5911 0.52094 Door 2 0.39276 0.35428 Door 3 0 0.12522 Door 4 0.30511 0.27513 Door 5 0.52632 0.46011 Door 6 0.30131 0.46892 Door 7 0.43305 0.38899 Door 8 0.45034 0.40642 Table C.63: Mixed strategies for reward structure 63 BRQR MATCH Door 1 0.45781 0.42106 Door 2 0.45493 0.43344 Door 3 0.37887 0.42106 Door 4 0.41701 0.38597 Door 5 0.46953 0.42983 Door 6 0 0.081882 Door 7 0.43749 0.42106 Door 8 0.38435 0.40571 Table C.64: Mixed strategies for reward structure 64 BRQR MATCH Door 1 0.45396 0.45077 Door 2 0.4555 0.41722 Door 3 0.46995 0.44006 Door 4 0.0056021 0.2936 Door 5 0.59619 0.43808 Door 6 0.26599 0.26453 Door 7 0.41032 0.37565 Door 8 0.34249 0.32009 Table C.65: Mixed strategies for reward structure 65 181 BRQR MATCH Door 1 0.39175 0.37822 Door 2 0.54356 0.35967 Door 3 0.29889 0.29712 Door 4 0.19516 0.27383 Door 5 0.30175 0.30129 Door 6 0.49048 0.51963 Door 7 0.32958 0.40198 Door 8 0.44882 0.46827 Table C.66: Mixed strategies for reward structure 66 BRQR MATCH Door 1 0.54516 0.49945 Door 2 0.069616 0.098137 Door 3 0.48612 0.45101 Door 4 0.48393 0.45756 Door 5 0.49695 0.46082 Door 6 0.40539 0.36744 Door 7 0.20291 0.29814 Door 8 0.30994 0.36744 Table C.67: Mixed strategies for reward structure 67 BRQR MATCH Door 1 0.37854 0.36109 Door 2 0.46088 0.43994 Door 3 0.29125 0.34719 Door 4 0.20145 0.24254 Door 5 0.45986 0.45136 Door 6 0.36941 0.36878 Door 7 0.42045 0.40916 Door 8 0.41816 0.37993 Table C.68: Mixed strategies for reward structure 68 182 BRQR MATCH Door 1 0.4422 0.36283 Door 2 0.55464 0.43804 Door 3 0.49131 0.44018 Door 4 0.24028 0.1803 Door 5 0.52049 0.44931 Door 6 0.45387 0.49177 Door 7 0.00014699 0.30973 Door 8 0.29706 0.32785 Table C.69: Mixed strategies for reward structure 69 BRQR MATCH Door 1 0.4041 0.39791 Door 2 0.53658 0.51833 Door 3 0.47882 0.44896 Door 4 0.40901 0.3819 Door 5 0.18271 0.22982 Door 6 0.36088 0.32299 Door 7 0.47326 0.46344 Door 8 0.15464 0.23666 Table C.70: Mixed strategies for reward structure 70 BRQR MATCH Door 1 0.44097 0.38996 Door 2 0.29028 0.33144 Door 3 0.61224 0.51286 Door 4 0.51033 0.4616 Door 5 0.15146 0.23937 Door 6 0.51439 0.48725 Door 7 0.14127 0.20988 Door 8 0.33906 0.36764 Table C.71: Mixed strategies for reward structure 71 183 BRQR MATCH Door 1 0.52781 0.54695 Door 2 0.43288 0.42504 Door 3 0.38384 0.33625 Door 4 0.3605 0.35533 Door 5 0.10065 0.18341 Door 6 0.28404 0.26142 Door 7 0.43291 0.43062 Door 8 0.47738 0.46098 Table C.72: Mixed strategies for reward structure 72 BRQR MATCH Door 1 0.30431 0.33387 Door 2 0.39066 0.38304 Door 3 0.24532 0.28946 Door 4 0.40868 0.3911 Door 5 0.3691 0.34473 Door 6 0.46677 0.4671 Door 7 0.46137 0.45711 Door 8 0.35378 0.33359 Table C.73: Mixed strategies for reward structure 73 BRQR MATCH Door 1 0.57141 0.53758 Door 2 0.10586 0.10494 Door 3 0.42112 0.53643 Door 4 0.38014 0.32019 Door 5 0.43545 0.38386 Door 6 0.25774 0.2403 Door 7 0.34384 0.35068 Door 8 0.48443 0.52602 Table C.74: Mixed strategies for reward structure 74 184 BRQR MATCH Door 1 0.32703 0.34091 Door 2 0.36262 0.31556 Door 3 0.46221 0.40014 Door 4 0.44843 0.39249 Door 5 0.34614 0.30019 Door 6 0.61849 0.53709 Door 7 0.27591 0.31325 Door 8 0.15916 0.40036 Table C.75: Mixed strategies for reward structure 75 BRQR MATCH Door 1 0.35316 0.31418 Door 2 0.53429 0.49219 Door 3 0.33999 0.32206 Door 4 0.4298 0.38325 Door 5 0.40742 0.39991 Door 6 0.45228 0.39989 Door 7 0.34306 0.28878 Door 8 0.14 0.39975 Table C.76: Mixed strategies for reward structure 76 BRQR MATCH Door 1 0.43721 0.40539 Door 2 0.39897 0.40539 Door 3 0.41051 0.35583 Door 4 0.40814 0.36345 Door 5 0.063105 0.25222 Door 6 0.46385 0.43364 Door 7 0.31285 0.305 Door 8 0.50536 0.47909 Table C.77: Mixed strategies for reward structure 77 185 BRQR MATCH Door 1 0.56592 0.53324 Door 2 0.28511 0.28561 Door 3 0.091549 0.26657 Door 4 0.47013 0.41663 Door 5 0.33376 0.36357 Door 6 0.44054 0.34993 Door 7 0.37489 0.3999 Door 8 0.43811 0.38456 Table C.78: Mixed strategies for reward structure 78 BRQR MATCH Door 1 0.41023 0.4141 Door 2 0.49533 0.45999 Door 3 0.11596 0.19237 Door 4 0.2946 0.28761 Door 5 0.40593 0.40265 Door 6 0.53505 0.47808 Door 7 0.38748 0.42999 Door 8 0.35542 0.33522 Table C.79: Mixed strategies for reward structure 79 BRQR MATCH Door 1 0.17808 0.35364 Door 2 0.41652 0.37862 Door 3 0.50074 0.45629 Door 4 0.43239 0.45629 Door 5 0.3001 0.24796 Door 6 0.51275 0.4634 Door 7 0.33493 0.33061 Door 8 0.3245 0.31321 Table C.80: Mixed strategies for reward structure 80 186 BRQR MATCH Door 1 0.14671 0.18259 Door 2 0.41762 0.39129 Door 3 0.63166 0.54911 Door 4 0.52872 0.49208 Door 5 0.24677 0.37792 Door 6 0.45888 0.43606 Door 7 0.26453 0.25399 Door 8 0.30511 0.31696 Table C.81: Mixed strategies for reward structure 81 BRQR MATCH Door 1 0.30979 0.30961 Door 2 0.47901 0.50961 Door 3 0.40944 0.36424 Door 4 0.44928 0.39972 Door 5 0.26792 0.28769 Door 6 0.38576 0.3901 Door 7 0.31966 0.32451 Door 8 0.37914 0.41452 Table C.82: Mixed strategies for reward structure 82 BRQR MATCH Door 1 0.48061 0.41884 Door 2 0.31624 0.26002 Door 3 0 0.28366 Door 4 0.21482 0.2943 Door 5 0.48607 0.38001 Door 6 0.44407 0.42738 Door 7 0.55418 0.4869 Door 8 0.50401 0.4489 Table C.83: Mixed strategies for reward structure 83 187 BRQR MATCH Door 1 0.47145 0.41113 Door 2 0.43511 0.39567 Door 3 0.41751 0.35219 Door 4 0.3277 0.31156 Door 5 0.24379 0.27336 Door 6 0.60648 0.55928 Door 7 0.17665 0.2412 Door 8 0.32131 0.45561 Table C.84: Mixed strategies for reward structure 84 BRQR MATCH Door 1 0.36963 0.34642 Door 2 0.084023 0.26146 Door 3 0.37142 0.33302 Door 4 0.54667 0.4689 Door 5 0.28798 0.42368 Door 6 0.45824 0.37156 Door 7 0.44427 0.39179 Door 8 0.43776 0.40318 Table C.85: Mixed strategies for reward structure 85 BRQR MATCH Door 1 0.48604 0.4257 Door 2 0.57309 0.51405 Door 3 0.47912 0.46033 Door 4 0.14453 0.46474 Door 5 0.12215 0.13141 Door 6 0.43166 0.35073 Door 7 0.41821 0.36371 Door 8 0.3452 0.28933 Table C.86: Mixed strategies for reward structure 86 188 BRQR MATCH Door 1 0 0 Door 2 0.42932 0.42063 Door 3 0.3466 0.37467 Door 4 0.25777 0.33245 Door 5 0.48519 0.458 Door 6 0.57703 0.5496 Door 7 0.45554 0.444 Door 8 0.44855 0.42063 Table C.87: Mixed strategies for reward structure 87 BRQR MATCH Door 1 0.27885 0.24926 Door 2 0.50591 0.48279 Door 3 0.32766 0.28862 Door 4 0.58417 0.52015 Door 5 0.224 0.21773 Door 6 0.32474 0.29254 Door 7 0.42292 0.4086 Door 8 0.33176 0.54031 Table C.88: Mixed strategies for reward structure 88 BRQR MATCH Door 1 0 0.22296 Door 2 0.47572 0.4627 Door 3 0.49254 0.40797 Door 4 0.49537 0.48455 Door 5 0.40848 0.3427 Door 6 0.57638 0.48034 Door 7 0.34615 0.42041 Door 8 0.20536 0.17837 Table C.89: Mixed strategies for reward structure 89 189 BRQR MATCH Door 1 0.44926 0.3942 Door 2 0.38558 0.39188 Door 3 0.37214 0.4855 Door 4 0.45608 0.37231 Door 5 0.44403 0.37681 Door 6 0.10908 0.18985 Door 7 0.16599 0.23308 Door 8 0.61784 0.55639 Table C.90: Mixed strategies for reward structure 90 BRQR MATCH Door 1 0.35937 0.34247 Door 2 0.46894 0.54012 Door 3 0.47832 0.45532 Door 4 0.22474 0.25343 Door 5 0.58586 0.50294 Door 6 0.42154 0.38598 Door 7 0 0.051061 Door 8 0.46123 0.46869 Table C.91: Mixed strategies for reward structure 91 BRQR MATCH Door 1 0.45596 0.40372 Door 2 0 0.23912 Door 3 0.31766 0.31757 Door 4 0.55963 0.51267 Door 5 0.4946 0.4768 Door 6 0.45358 0.39691 Door 7 0.33961 0.32083 Door 8 0.37895 0.33239 Table C.92: Mixed strategies for reward structure 92 190 BRQR MATCH Door 1 0.30274 0.27302 Door 2 0.46241 0.39182 Door 3 0.39765 0.41249 Door 4 0.51947 0.46874 Door 5 0.55123 0.52749 Door 6 0.024132 0.22767 Door 7 0.44453 0.3993 Door 8 0.29784 0.29947 Table C.93: Mixed strategies for reward structure 93 BRQR MATCH Door 1 0.58607 0.55635 Door 2 0.38565 0.38453 Door 3 0.25527 0.33504 Door 4 0.38271 0.3717 Door 5 0.060333 0.040625 Door 6 0.66425 0.61211 Door 7 0.28311 0.29316 Door 8 0.38262 0.40647 Table C.94: Mixed strategies for reward structure 94 BRQR MATCH Door 1 0.44436 0.4128 Door 2 0.050399 0.21401 Door 3 0.5048 0.44707 Door 4 0.47632 0.45977 Door 5 0.51961 0.46532 Door 6 0.45607 0.42472 Door 7 0.33367 0.33977 Door 8 0.21478 0.23654 Table C.95: Mixed strategies for reward structure 95 191 BRQR MATCH Door 1 0.45507 0.37696 Door 2 0.3195 0.29804 Door 3 0.2552 0.30235 Door 4 0.41958 0.46596 Door 5 0.43017 0.42488 Door 6 0.42757 0.41123 Door 7 0.3291 0.36823 Door 8 0.36381 0.35235 Table C.96: Mixed strategies for reward structure 96 BRQR MATCH Door 1 0.50416 0.45682 Door 2 0.28621 0.2567 Door 3 0.40689 0.41981 Door 4 0.39385 0.40644 Door 5 0.11433 0.16925 Door 6 0.44542 0.45322 Door 7 0.53501 0.52515 Door 8 0.31413 0.31261 Table C.97: Mixed strategies for reward structure 97 BRQR MATCH Door 1 0.21249 0.16662 Door 2 0.48702 0.38328 Door 3 0.52795 0.4848 Door 4 0.15425 0.51505 Door 5 0.3449 0.38328 Door 6 0.49103 0.42705 Door 7 0.39813 0.3333 Door 8 0.38422 0.30662 Table C.98: Mixed strategies for reward structure 98 192 BRQR MATCH Door 1 0.45856 0.42273 Door 2 0.20426 0.19968 Door 3 0.31772 0.3236 Door 4 0.51708 0.41883 Door 5 0.59705 0.56818 Door 6 0.33064 0.37753 Door 7 0.084073 0.19968 Door 8 0.49063 0.48978 Table C.99: Mixed strategies for reward structure 99 BRQR MATCH Door 1 0.44257 0.44742 Door 2 0.56432 0.51979 Door 3 0.18584 0.26649 Door 4 0.29319 0.27808 Door 5 0.25455 0.25583 Door 6 0.29668 0.30456 Door 7 0.64627 0.64974 Door 8 0.31657 0.27808 Table C.100: Mixed strategies for reward structure 100 BRQR MATCH Door 1 0.49679 0.4084 Door 2 0.20353 0.23884 Door 3 0.4179 0.38812 Door 4 0.52119 0.4605 Door 5 0.24172 0.25875 Door 6 0.22153 0.38272 Door 7 0.38888 0.42411 Door 8 0.50847 0.43857 Table C.101: Mixed strategies for reward structure 101 193 BRQR MATCH Door 1 0.19414 0.16762 Door 2 0.16265 0.2379 Door 3 0.65623 0.6208 Door 4 0.57062 0.56333 Door 5 0.43565 0.40462 Door 6 0.01059 0.040001 Door 7 0.41914 0.38727 Door 8 0.55098 0.57846 Table C.102: Mixed strategies for reward structure 102 BRQR MATCH Door 1 0.40715 0.35545 Door 2 0.53811 0.47994 Door 3 0.2162 0.33738 Door 4 0.55561 0.49557 Door 5 0.64485 0.56545 Door 6 0 0.044228 Door 7 0.11716 0.2699 Door 8 0.52093 0.45209 Table C.103: Mixed strategies for reward structure 103 BRQR MATCH Door 1 0.25257 0.28932 Door 2 0.35074 0.34951 Door 3 0.24863 0.25359 Door 4 0.42363 0.42453 Door 5 0.41793 0.43614 Door 6 0.36486 0.34592 Door 7 0.48558 0.46484 Door 8 0.45606 0.43614 Table C.104: Mixed strategies for reward structure 104 194 BRQR MATCH Door 1 0.24216 0.25119 Door 2 0.37358 0.35052 Door 3 0.34624 0.31754 Door 4 0.46241 0.4301 Door 5 0.48172 0.44378 Door 6 0.15178 0.3602 Door 7 0.6298 0.56648 Door 8 0.31232 0.28018 Table C.105: Mixed strategies for reward structure 105 BRQR MATCH Door 1 0.21946 0.34945 Door 2 0.38255 0.41575 Door 3 0.40698 0.3523 Door 4 0.32616 0.30727 Door 5 0.65287 0.58288 Door 6 0.47732 0.39963 Door 7 0 0.12431 Door 8 0.53467 0.46841 Table C.106: Mixed strategies for reward structure 106 BRQR MATCH Door 1 0.26473 0.36808 Door 2 0.45255 0.40867 Door 3 0.31198 0.26788 Door 4 0.51445 0.44713 Door 5 0.55379 0.44713 Door 6 0.53651 0.45835 Door 7 0.26508 0.20141 Door 8 0.10091 0.40134 Table C.107: Mixed strategies for reward structure 107 195 BRQR MATCH Door 1 0.38035 0.34287 Door 2 0.45712 0.41989 Door 3 0.19809 0.20319 Door 4 0.25018 0.34807 Door 5 0.3774 0.36073 Door 6 0.35367 0.29452 Door 7 0.57872 0.66609 Door 8 0.40447 0.36464 Table C.108: Mixed strategies for reward structure 108 196 Appendix D: Expected Rewards for COBRA Experiments DOBSS COBRA COBRA COBRA COBRA MAXIMIN UNIFORM (0,2.5) (1,2.5) (.37,2.5) (.03,2.5) Door 1 -5 -4.58 -5 -5 -4.60 -1.62 -2.78 Door 2 -.96 -.42 1.35 -.29 -.36 -1.62 -3.56 Door 3 .35 -.35 -1 -.29 -.36 -1 .11 Door 4 -1.37 -.79 1.35 -.29 -.73 -1.62 -2.67 Door 5 .05 -.35 1.35 -.29 -.36 -1.62 -1.67 Door 6 -1 -.86 -1 -1 -.86 -1 -.26 Door 7 .38 -.35 1.35 -.29 -.36 -1.62 -2.56 Door 8 -4.56 -3.68 -7 -3.32 -3.67 -1.62 -3.67 Table D.1: Expected rewards for reward structure 1 DOBSS COBRA COBRA COBRA COBRA MAXIMIN UNIFORM (0,2.5) (1,2.5) (.54,2.5) (.41,2.5) Door 1 -1.32 -1.09 4 .09 -.57 -1.63 -3.56 Door 2 -4.21 -2.81 -10 -2.12 -2.69 -1.63 -5.19 Door 3 -.62 -1 -1 -.89 -.72 -1 -.26 Door 4 .79 -1.09 5 .09 -.57 -1.63 -3.19 Door 5 -1 -1 -1 -1 -1 -1 -.26 Door 6 -2.06 -1.44 -3 -2.72 -2.31 -1.63 -1.15 Door 7 -.64 -1.09 5 .09 -.57 -1.63 -5.08 Door 8 -2.88 -2.12 -5 -2.93 -2.75 -1.63 -2.41 Table D.2: Expected rewards for reward structure 2 197 DOBSS COBRA COBRA COBRA COBRA MAXIMIN UNIFORM (0,2.5) (1,2.5) (.75,2.5) (.25,2.5) Door 1 -.92 -1.66 4 .62 -1.31 -2.12 -3.56 Door 2 -1.51 -1.66 -5 -.16 -1.31 -2.12 -2.04 Door 3 -.80 -1 -1 -1 -1 -1 -.26 Door 4 -.21 -1.66 5 .62 -1.31 -2.12 -4.45 Door 5 -5 -3.92 -5 -5 -5 -2.12 -2.78 Door 6 -1.76 -1.66 -3 -3 -1.31 -2.12 -1.15 Door 7 -.26 -1.66 5 .62 -1.31 -2.12 -3.82 Door 8 -3.16 -1.74 -6 -3.75 -1.87 -2.12 -3.04 Table D.3: Expected rewards for reward structure 3 DOBSS COBRA COBRA COBRA COBRA MAXIMIN UNIFORM (0,2.5) (1,2.5) (.75,2.5) (.25,2.5) Door 1 -1.51 -1.51 3 -.50 -1.51 -1.51 -3.93 Door 2 -1.51 -1.51 -5 -.50 -1.51 -1.51 -2.04 Door 3 -1.51 -1.51 -2 -2 -1.51 -1.51 -.15 Door 4 -1.51 -1.51 3 -.50 -1.51 -1.51 -5.19 Door 5 -1 -1 -1 -1 -1 -1 .48 Door 6 -1.51 -1.51 -3 -3 -1.51 -1.51 -.78 Door 7 -1.51 -1.51 3 -.50 -1.51 -1.51 -4.56 Door 8 -1.51 -1.51 -4 -1.75 -1.51 -1.51 -1.41 Table D.4: Expected rewards for reward structure 4 198 Appendix E: Expected Response Percentages for COBRA Experiment Structure One Unobserved 5 20 Unlimited DOBSS 20% 7.5% 17.5% 12.5% COBRA(0,") 65% 65% 65% 70% COBRA(,") 57.5% 92.5% 72.5% 70% COBRA(C,") 92.5% 92.5% 87.5% 95% MAXIMIN 100% 100% 100% 100% Structure Two DOBSS 27.5% 25% 12.5% 10% COBRA(0,") 62.5% 65% 40% 55% COBRA(,") 62.5% 57.5% 47.5% 55% COBRA(C,") 62.5% 57.5% 55% 47.5% MAXIMIN 100% 100% 100% 100% Structure Three DOBSS 20% 20% 20% 25% COBRA(0,") 75% 72.5% 62.5% 60% COBRA(,") 50% 47.5% 67.5% 60% COBRA(C,") 50% 47.5% 20% 25% MAXIMIN 100% 100% 100% 100% Structure Four DOBSS 100% 92.5% 87.5% 85% COBRA(0,") 100% 92.5% 87.5% 85% COBRA(," 42.5% 52.5% 82.5% 85% COBRA(C,") 52.5% 52.5% 25% 35% MAXIMIN 100% 100% 100% 100% Table E.1: Percentage of times follower chose an expected strategy 199 Appendix F: Strategies for varying in COBRA(,2.5) Door 1 Door 2 Door 3 Door 4 Door 5 Door 6 Door 7 Door 8 =:00 .069 .631 .213 .578 .515 .069 .553 .368 =:05 .062 .635 .208 .592 .513 .062 .552 .372 =:10 .056 .634 .203 .608 .512 .056 .550 .377 =:15 .051 .632 .198 .621 .510 .051 .549 .384 =:20 .050 .632 .194 .620 .509 .050 .548 .394 =:25 .049 .630 .190 .619 .507 .049 .547 .406 =:30 .046 .630 .187 .618 .506 .046 .546 .418 =:35 .005 .640 .227 .631 .520 .005 .556 .414 =:40 .000 .643 .242 .636 .525 .000 .560 .391 =:45 .000 .647 .256 .641 .529 .000 .564 .361 =:50 .000 .651 .272 .646 .535 .000 .568 .325 =:55 .000 .656 .292 .653 .542 .000 .573 .282 =:60 .000 .662 .318 .661 .550 .000 .579 .227 =:65 .000 .671 .350 .672 .561 .000 .587 .156 =:70 .000 .681 .393 .686 .575 .000 .598 .063 =:75 .000 .689 .423 .696 .585 .000 .605 .000 =:80 .000 .689 .423 .696 .585 .000 .605 .000 =:85 .000 .689 .423 .696 .585 .000 .605 .000 =:90 .000 .689 .423 .696 .585 .000 .605 .000 =:95 .000 .731 .227 .752 .641 .000 .647 .000 = 1:00 .000 .779 .000 .817 .706 .000 .696 .000 Table F.1: -variations for reward structure 1 200 Door 1 Door 2 Door 3 Door 4 Door 5 Door 6 Door 7 Door 8 =:00 .575 .553 .000 .530 .000 .311 .618 .410 =:05 .579 .555 .000 .535 .000 .300 .622 .405 =:10 .584 .557 .006 .539 .000 .287 .625 .399 =:15 .587 .557 .026 .542 .000 .268 .628 .389 =:20 .591 .556 .048 .546 .000 .248 .631 .377 =:25 .595 .555 .074 .549 .000 .224 .634 .365 =:30 .600 .554 .103 .554 .000 .198 .637 .350 =:35 .606 .554 .136 .559 .000 .167 .642 .334 =:40 .617 .560 .139 .569 .000 .139 .650 .322 =:45 .634 .573 .114 .585 .000 .114 .663 .314 =:50 .654 .589 .084 .604 .000 .084 .678 .304 =:55 .679 .609 .046 .627 .000 .046 .697 .292 =:60 .711 .634 .000 .656 .000 .000 .720 .277 =:65 .730 .633 .000 .674 .000 .000 .735 .225 =:70 .757 .632 .000 .698 .000 .000 .755 .156 =:75 .793 .631 .000 .732 .000 .000 .782 .059 =:80 .828 .597 .000 .765 .000 .000 .809 .000 =:85 .863 .503 .000 .797 .000 .000 .635 .000 =:90 .933 .316 .000 .861 .000 .000 .887 .000 =:95 1.00 .000 .000 1.00 .062 .000 .937 .000 = 1:00 1.00 .000 .000 1.00 .000 .000 1.00 .000 Table F.2: -variations for reward structure 2 201 Door 1 Door 2 Door 3 Door 4 Door 5 Door 6 Door 7 Door 8 =:00 .527 .416 .000 .555 .179 .266 .523 .531 =:05 .532 .423 .000 .559 .151 .277 .527 .529 =:10 .537 .431 .000 .563 .119 .290 .532 .526 =:15 .543 .440 .000 .568 .083 .304 .537 .523 =:20 .550 .450 .000 .573 .043 .320 .542 .520 =:25 .557 .460 .000 .579 .000 .337 .549 .516 =:30 .559 .464 .000 .580 .000 .342 .550 .502 =:35 .562 .468 .000 .583 .000 .345 .553 .487 =:40 .570 .481 .000 .590 .000 .320 .560 .476 =:45 .581 .496 .000 .598 .000 .290 .569 .464 =:50 .593 .514 .000 .607 .000 .254 .579 .450 =:55 .607 .536 .000 .619 .000 .210 .592 .432 =:60 .626 .564 .000 .634 .000 .155 .608 .410 =:65 .650 .600 .000 .653 .000 .085 .628 .381 =:70 .687 .615 .000 .683 .000 .001 .660 .352 =:75 .719 .604 .000 .708 .000 .000 .687 .280 =:80 .766 .587 .000 .746 .000 .000 .728 .171 =:85 .842 .556 .000 .807 .000 .000 .293 .000 =:90 .908 .381 .000 .860 .000 .000 .850 .000 =:95 1.00 .071 .000 1.00 .000 .000 .928 .000 = 1:00 1.00 .000 .000 1.00 .000 .000 1.00 .000 Table F.3: -variations for reward structure 3 202 Door 1 Door 2 Door 3 Door 4 Door 5 Door 6 Door 7 Door 8 =:00 .589 .435 .096 .652 .000 .247 .623 .354 =:05 .589 .435 .096 .652 .000 .247 .623 .354 =:10 .589 .435 .096 .652 .000 .247 .623 .354 =:15 .589 .435 .096 .652 .000 .247 .623 .354 =:20 .589 .435 .096 .652 .000 .247 .623 .354 =:25 .589 .435 .096 .652 .000 .247 .623 .354 =:30 .589 .435 .096 .652 .000 .247 .623 .354 =:35 .589 .435 .096 .652 .000 .247 .623 .354 =:40 .589 .435 .096 .652 .000 .247 .623 .354 =:45 .589 .435 .096 .652 .000 .247 .623 .354 =:50 .589 .435 .096 .652 .000 .247 .623 .354 =:55 .600 .451 .010 .662 .000 .268 .634 .372 =:60 .609 .462 .000 .669 .000 .231 .641 .285 =:65 .624 .482 .000 .681 .000 .146 .655 .409 =:70 .651 .520 .000 .705 .000 .048 .680 .393 =:75 .681 .561 .000 .730 .000 .000 .707 .320 =:80 .712 .604 .000 .756 .000 .000 .736 .190 =:85 .787 .477 .000 .819 .000 .000 .804 .010 =:90 .852 .406 .000 .875 .000 .000 .865 .000 =:95 1.00 .000 .000 1.00 .000 .000 1.00 .000 = 1:00 1.00 .000 .000 1.00 .000 .000 1.00 .000 Table F.4: -variations for reward structure 4 203 Appendix G: Experimental Instructions G.1 Material for COBRA Experiments You will play 14 unique and distinct instances of the following game. Please note that each instance is neither related to nor involves any of the other instances of the game. No memory will be maintained of either the pirates’ actions or your choices. We will describe below how one instance of the game unfolds. Remember you will play 14 instances of the game. Pirate Game: *These may not be the rewards you will see Figure G.1: Game Interface Description: As you were boating around one day you stumbled across an island inhabited by pirates and quickly discovered that they were storing mass amounts of treasure on the island. Knowing that this treasure was stolen from hard working people you thought it would only be fair to try and get some of this hard earned money back from the pirates. After some observation over several days you have found the following: Layout: There are eight doors in total which lead to the pirates’ treasure. 204 Figure G.2: Single Observation At any given time three of the eight doors will be guarded by pirates as shown in Figure Two. Each of the eight doors has a dierent amount of treasure behind it and thus is worth dierent values to both you and the pirates. The rewards for each of you are given as shown in both figures*. Both you and the pirate are aware of each others rewards and penalties. If you succeed then you get your reward and the pirate incurs their penalty. If you are caught then you incur your penalty and the pirate receives their reward. So initially, you see a picture as shown in Figure 1, where you don’t know the pirates’ position. When you make your choice, the pirates’ positions will be revealed. In the above example, if you chose door 0 to steal from, you will get caught, incurring a penalty of -3, while the pirate wins a reward of 4. If you had instead chosen to steal from door 1, you would have obtained a reward of 5. You can only attempt to steal from a single door. Once you have made your choice you will be informed whether you got caught or not. Rules: You will be given a specific number of observations to attempt and learn the pirates’ strat- egy. Each observation consists of the doors opening to show you the locations of the pirates (i.e. which doors the pirates are guarding), and the pirates may change which doors they guard from day to day. Once your observations are complete you will be given an unlimited amount of time to choose a door to try and steal from. In essence, having observed the pirates for some number of days, you now have to make your move and steal from a particular door. 205 Once a door is chosen you will see whether you got away with the treasure or got caught and the appropriate points will either be added or deducted from your overall score. Conclusion: At the end of the twelve individual games your total score will decide how much money you earn. Each point you earn is worth 15 cents. The pirates’ points have no bearing on how much money you will get. You will begin the game with 8 free dollars. If you have a negative score you will lose 15 cents per negative point up to $3. You will be guaranteed to leave with at least $5 no matter how negative your score is. If you have any questions feel free to ask 206 G.2 Material for MATCH Experiments G.2.1 Obvious Games Gates 1 2 3 4 5 6 7 8 Subject Reward 10 1 9 2 3 10 2 4 Subject Penalty -1 -5 -9 -2 -5 -8 -5 -3 Defender Reward 5 6 2 8 4 2 1 4 Defender Penalty -2 -3 -2 -2 -3 -3 -3 -2 Defender Strategy .05 .30 .50 .35 .30 .65 .35 .50 Table G.1: Dummy structure 1 Gates 1 2 3 4 5 6 7 8 Subject Reward 2 9 4 10 3 2 5 10 Subject Penalty -5 -8 -9 -10 -5 -8 -5 -1 Defender Reward 5 6 2 8 4 2 1 4 Defender Penalty -2 -3 -2 -2 -3 -3 -3 -2 Defender Strategy .30 .70 .25 .60 .25 .35 .50 .05 Table G.2: Dummy structure 2 207 G.3 Experiment Instructions Instructions How is the game played? 208 209 210 211 212 213
Abstract (if available)
Linked assets
University of Southern California Dissertations and Theses
Conceptually similar
PDF
Human adversaries in security games: integrating models of bounded rationality and fast algorithms
PDF
Balancing tradeoffs in security games: handling defenders and adversaries with multiple objectives
PDF
Predicting and planning against real-world adversaries: an end-to-end pipeline to combat illegal wildlife poachers on a global scale
PDF
Thwarting adversaries with unpredictability: massive-scale game-theoretic algorithms for real-world security deployments
PDF
Keep the adversary guessing: agent security by policy randomization
PDF
Modeling human bounded rationality in opportunistic security games
PDF
Addressing uncertainty in Stackelberg games for security: models and algorithms
PDF
Game theoretic deception and threat screening for cyber security
PDF
The interpersonal effect of emotion in decision-making and social dilemmas
PDF
Towards addressing spatio-temporal aspects in security games
PDF
Protecting networks against diffusive attacks: game-theoretic resource allocation for contagion mitigation
PDF
Hierarchical planning in security games: a game theoretic approach to strategic, tactical and operational decision making
PDF
Combating adversaries under uncertainties in real-world security problems: advanced game-theoretic behavioral models and robust algorithms
PDF
When AI helps wildlife conservation: learning adversary behavior in green security games
PDF
Context dependent utility: an appraisal-based approach to modeling context, framing, and decisions
PDF
Handling attacker’s preference in security domains: robust optimization and learning approaches
PDF
Not a Lone Ranger: unleashing defender teamwork in security games
PDF
Real-world evaluation and deployment of wildlife crime prediction models
PDF
Computational model of human behavior in security games with varying number of targets
PDF
Automated negotiation with humans
Asset Metadata
Creator
Pita, James A.
(author)
Core Title
The human element: addressing human adversaries in security domains
School
Viterbi School of Engineering
Degree
Doctor of Philosophy
Degree Program
Computer Science
Publication Date
12/04/2012
Defense Date
12/15/2012
Publisher
University of Southern California
(original),
University of Southern California. Libraries
(digital)
Tag
behavioral game theory,Decision making,game theory,OAI-PMH Harvest,robust decision making,Security
Language
English
Contributor
Electronically uploaded by the author
(provenance)
Advisor
Tambe, Milind (
committee chair
), Gratch, Jonathan (
committee member
), John, Richard S. (
committee member
), Kraus, Sarit (
committee member
), Marsella, Stacy C. (
committee member
), Weller, Nicholas (
committee member
)
Creator Email
jpita@usc.edu,pitajames@gmail.com
Permanent Link (DOI)
https://doi.org/10.25549/usctheses-c3-125877
Unique identifier
UC11291680
Identifier
usctheses-c3-125877 (legacy record id)
Legacy Identifier
etd-PitaJamesA-1380.pdf
Dmrecord
125877
Document Type
Dissertation
Rights
Pita, James A.
Type
texts
Source
University of Southern California
(contributing entity),
University of Southern California Dissertations and Theses
(collection)
Access Conditions
The author retains rights to his/her dissertation, thesis or other graduate work according to U.S. copyright law. Electronic access is being provided by the USC Libraries in agreement with the a...
Repository Name
University of Southern California Digital Library
Repository Location
USC Digital Library, University of Southern California, University Park Campus MC 2810, 3434 South Grand Avenue, 2nd Floor, Los Angeles, California 90089-2810, USA
Tags
behavioral game theory
game theory
robust decision making