Dynamics of Multi-Agent Learning Under Bounded Rationality: Theory and Empirical Evidence

Abstract

This thesis contributes to the development of a principled understanding of the learning dynamics and strategic interactions in human-machine systems. We propose a game-theoretic framework that captures the complexities of multi-agent learning under bounded rationality, focusing on the effects of timescale separation, varying cost structures, and the ability to anticipate each other's reactions. By leveraging tools from continuous games, dynamical systems, and control theory, we characterize the stability and convergence properties of learning dynamics in multi-agent settings, providing insights into leader-follower structures, consistent conjectures, and behavior shaping. We validate our theoretical findings through a series of human-machine experiments, demonstrating the practical implications of our approach for the design and control of machine learning systems that interact with humans. Our work highlights the importance of considering the ethical implications of advanced AI systems and emphasizes the need for developing AI alignment solutions and cognitive science research to ensure that these systems are designed to be robust, beneficial, and aligned with human values. The proposed framework and empirical findings contribute to the scientific understanding of strategic reasoning, adaptation, and decision-making in human-machine systems, laying the foundation for the responsible development and deployment of adaptive technologies in real-world applications.