CAREER: Learning Agents in Dynamic, Collaborative, and Adversarial Multiagent Environments

University Of Texas At Austin, Austin TX

Investigators

Abstract

This project aims to enable multiple intelligent agents to learn to act both individually and in coordination with one another towards individual and/or common goals in real-time, noisy, collaborative and adversarial environments. The approach taken will be to study complete agents in specific, complex environments, with the goal of drawing general lessons from the specific implementations. Fundamental research will be conducted in four main areas. First, multiagent reinforcement learning will be scaled up to handle larger and more complex problems than has been previously possible. Second, new state representations suitable for learning will be proposed and tested. Third, game theoretic approaches to improving agent performance by predicting the responses of other agents will be investigated. Fourth, strategies for learning autonomous bidding agents will be developed and tested. Application domains will include: robotic soccer, both in simulation and with real robots; and autonomous bidding agents in multiple realistic scenarios. The rich simulation environments to be used for this research are ideal substrates for teaching students about complete intelligent agents, including perception, cognition, and action. The educational goals of this project include leveraging the appeal of these domains to students into challenging, exciting, and instructive undergraduate and graduate courses.