Foundations of Multiagent Reinforcement Learning
Abstract
Project AbstractApproved for Public Release.Many of the most exciting recent applications of reinforcement learning---including Go,,Poker, real-time strategy games, decentralized controls, autonomous driving---are game-theoretic in nature. Agents must learn in the, presence of other agents whose decisions can adapt to the strategies of the agents. While the single-agent reinforcement learning p,roblem has been the subject of intense recent investigation---including development of efficient algorithms with provable, non-asymp,totic theoretical guarantees---Multi-agent Reinforcement Learning (MARL) has been comparatively unexplored. This proposal aims to es,tablish a deep and solid theoretical foundation for MARL. The project will investigate MARL under the model of Markov Games (MGs, Sh,apley 1953), and provide a comprehensive set of theoretical results on the optimal complexity of learning various equilibria in the,settings of (1) two-player zero-sum MGs, (2) multiplayer general-sum MGs, and (3) MGs with a enormously large state space. To accomp,lish this, this project will identify and address fundamental challenges that are unique in MARL, and design new classes of highly e,fficient algorithms which reliably find the desired equilibria using a small number of samples.Impact on DOD and Navy. The proposed,topics are of fundamental importance in enabling real-time, automated, tactical decision making in the presence of multiple intellig,ent agents. This project will help the future development of reliable and highly effective AI-assist decision making systems to supp,ort the tactics and operations in modern complex ocean battlespace.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Apr 01, 2022
- Source ID
- N000142212253
Entities
People
- Chi Jin
Organizations
- Office of Naval Research
- Trustees of Princeton University
- United States Navy