Decision-Theoretic Planning in Multi-Agent Environments

University Of Illinois At Chicago, Chicago IL

Investigators

Abstract

This work will develop interactive partially observable Markov decision processes (IPOMDPs), which provide a framework for optimal planning that agents can use while interacting with other agents. I-POMDPs unify decision-theoretic approach to planning with elements of games with incomplete information. Decision-theoretic planning, formalized as traditional POMDPs, has had a wide impact in control theory, artificial intelligence, and cognitive science. However, POMDPs by themselves are not suitable in interactive settings because they would force the assumption that other agents are random and static. Game theory, on the other hand, is centered on interaction among rational agents. However, the game-theoretic solution concept of Nash equilibria and their variations result in solutions that are non-unique and incomplete. I-POMDPs address the above disconnect between decision and game theories. They include Bayesian update over models of other agents during interactions, result in policies that are best responses to the others' expected actions, and reduce to classical POMDPs if the agent is acting alone.