Graduation Semester and Year
Spring 2025
Language
English
Document Type
Thesis
Degree Name
Master of Science in Computer Science
Department
Computer Science and Engineering
First Advisor
Dr. Manfred Huber
Second Advisor
Dr. Farhad Kamangar
Third Advisor
Dr. David Levine
Abstract
This thesis explores the emergence of trust, deception, and adaptive strategy in multi-agent reinforcement learning (MARL) environments using the social deduction game Werewolf as a simulation framework. In this environment, agents operate with hidden roles, incomplete information, and the need to reason about others’ intentions- mirroring the complexities of real-world social interactions. We present and evaluate two agent architectures: Agent vA, a symbolic, heuristic-based agent with probabilistic trust modeling and scalable memory structures; and Agent vB, a modular Q-learning agent that learns phase-specific policies through reinforcement. Agent vA relies on symbolic reasoning, bounded belief updates, and generalizable heuristics, while Agent vB uses a finely-grained state-action space for learning detailed strategies. Through controlled experiments across varying population sizes and role distributions, we find that Agent vA demonstrates better scalability, robustness to environmental shifts, and early-game strategic advantage. In contrast, Agent vB learns its behavior entirely through reinforcement, enabling the discovery of nuanced strategies tailored to each game phase—night actions, symbolic communication, and voting. Its modular Q learning architecture allows for role-specific policy formation and rapid exploitation of recurring patterns, often leading to efficient convergence and strong performance in fixed environments. Our results show that while reinforcement learning enables emergent deception and collaboration, symbolic agents offer superior flexibility and performance in dynamic or previously unseen environments. These findings provide insight into designing robust MARL agents for adversarial and socially complex domains, with implications for trustworthy AI, strategic planning, and cognitive modeling.
Keywords
(Multi-agent reinforcement learning, Trust modeling, Strategic deception, Social deduction games, Q-learning, Werewolf game simulation, Agent-based modeling)
Disciplines
Artificial Intelligence and Robotics | Cognitive Science
License
This work is licensed under a Creative Commons Attribution 4.0 International License.
Recommended Citation
Patel, Pathikkumar Dharmeshbhai, "Modeling Trust and Deception in Multi-Agent Reinforcement Learning Using the Werewolf Game" (2025). Computer Science and Engineering Theses. 529.
https://mavmatrix.uta.edu/cse_theses/529
Comments
First and foremost, I would like to express my deepest gratitude to my advisor, Dr. Manfred Huber. His extraordinary patience, unwavering support, and genuine passion for teaching have been a guiding light throughout my academic journey. No matter how confused or lost I felt, Dr. Huber always responded with calm clarity, kindness, and encouragement. His ability to mentor without judgment and to teach with humility has not only shaped my research but also left a lasting impact on me personally. I consider myself incredibly fortunate to have had the opportunity to learn under his guidance. I would also like to sincerely thank the members of my thesis committee for their valuable feedback and thoughtful suggestions, which have been instrumental in shaping this work. Mydeepest appreciation goes to my family—my mother, father, and grandmother—for their unconditional love, sacrifices, and belief in me. Their support has been the foun dation of all that I’ve achieved. I also carry the memory of my late grandfather, whose quiet strength and values continue to guide me—his presence remains with me in spirit every step of the way. A heartfelt thanks to my extended family in Waco, who welcomed me with open arms and made me feel at home when I was far from it. Their warmth and genuine care turned an unfamiliar place into a second home. From home-cooked meals and quiet support to simply being there when I needed a break, they gave me the comfort and stability to stay focused and at peace. Finally, to all my friends and everyone who stood by me—whether through late-night study sessions, shared laughter, or simple encouragement—thank you for being a part of this journey