Improving Ad-Hoc Cooperation in Multiagent Reinforcement Learning via Skill Modeling

Abstract

Machine learning is a versatile tool allowing for, among other things, training intelligent agents capable of autonomously acting in their environments. In particular, Multiagent Reinforcement Learning has made tremendous progress enabling such agents to interact with one another in an effective manner. One of the challenges that this field is still facing, however, is the problem of ad-hoc cooperation, or cooperation with agents that have not been previously encountered.

This thesis explores one possible approach to tackle this issue, using the psychology-inspired idea of Theory of Mind. Specifically, a component designed to explicitly model the skill level of the other agent is included, to allow the primary agent to better choose its actions.

The results show that this approach does in fact facilitate better coordination in an environment designed to test this skill and is a promising method for more complicated scenarios.

The potential applications can be found in any situation that requires coordination between multiple intelligent agents (which may also include humans), such as traffic coordination between autonomous vehicles, or rescue operations where autonomous agents and humans have to work together to efficiently search an area.