Transfer in the Context of Reinforcement Learning by Mapping Q-Tables

Transfer in machine learning is the process of using knowledge learned in a source domain to speed learning in one or more related target domains. In human learning, transfer is a ubiquitous phenomenon. In machine learning, transfer is far less common. In this thesis we present a transfer method for reinforcement learning when the learner does not have access to a model of either the source or the target domains (i.e., the transition and reward probabilities are unknown) and there is no prior knowledge about how to map states or actions in the source domain to corresponding or similar states or actions in the target domain. Empirical results in a variety of grid worlds and a multi-agent block loading domain that is exceptionally difficult to solve using standard reinforcement learning algorithms show significant speedups in learning in the target domain.