Peer Reinforcement in Homogeneous and Heterogeneous Multi-agent Learning

摘要

Reinforcement learning is a broadly employed methodology for training adaptive agents in single- and multi-agent settings. Existing approaches, while being able to vary the type and nature of reinforcement, rely heavily on a centralized omniscient source for reinforcement. This is a significant limitation in terms of modelling human learning: while we do learn directly from skilled teachers, we also learn much from those around us participating in the group activities. Ignoring the latter source of reinforcement severely limits the amount of information an agent can obtain from the world around it. In this paper we explore the use of peer reinforcement – reinforcement obtained from others participating in the same activity, and the effects of employing peer reinforcement for learning in multi-agent systems. We examine two scenarios in a robotic soccer domain to illustrate the use of peer reinforcement in both heterogeneous and homogeneous multi-agent settings.