Transfer in Reinforcement Learning via Shared Features Citation

We present a framework for transfer in reinforcement learni ng based on the idea that related tasks share some common features, and that transfer can be achieve d via those shared features. The framework attempts to capture the notion of tasks that are re lated but distinct, and provides some insight into when transfer can be usefully applied to a probl em sequence and when it cannot. We apply the framework to the knowledge transfer problem, and s how that an agent can learn a portable shaping function from experience in a sequence of tasks to si gnificantly improve performance in a later related task, even given a very brief training period . We also apply the framework to skill transfer, to show that agents can learn portable skills acro ss a sequence of tasks that significantly improve performance on later related tasks, approaching th e performance of agents given perfectly learned problem-specific skills.