Reinforcement Learning with Imitation in Heterogeneous Multi-Agent Systems

The application of decision making and learning algorithms to multi-agent systems presents many interestingresearch challenges and opportunities. Among these is the ability for agents to learn how to act by observing or imitating other agents. We describe an algorithm, the IQ-algorithm, that integrates imitation with Q-learning. Roughly, a Q-learner uses the observations it has made of an “expert” agent to bias its exploration in promising directions. This algorithm goes beyond previous work in this direction by relaxing the oft-made assumptions that the learner (observer) and the expert (observed agent) share the same objectives and abilities. Our preliminary experiments demonstrate significant transfer between agents using the IQ-model and in many cases reductions in training time.