Speaker
Antonios Varvitsiotis
Singapore University of Technology
Topic
Identifying and controlling agent behavior in games using limited data
Decentralized learning algorithms are an essential tool for designing multi-agent systems, as they enable agents to autonomously learn from their experience and past interactions. In this work, we propose a theoretical and algorithmic framework for real-time identification of the learning dynamics that govern agent behavior in games using a short burst of a single trajectory. Our method identifies agent dynamics through polynomial regression, where we compensate for limited data by incorporating side-information constraints that capture fundamental assumptions or expectations about agent behavior, e.g., agents tend to move towards directions of improving utility. These constraints are enforced computationally using sum-of-squares optimization, leading to a hierarchy of increasingly better approximations of the true agent dynamics. Extensive experiments demonstrate that our approach accurately recovers the true dynamics across various games and target learning dynamics while using only five samples from a short run of a single trajectory. Notably, we use strong benchmarks such as predicting equilibrium selection as well as the evolution of chaotic systems for up to ten Lyapunov times.
Zoom-Link for online participation: https://cwi-nl.zoom.us/j/87542191297?pwd=QTIxYlNQUmQ5UTFKb0NSaVVsbGFJZz09
