Task similarity measures for transfer in reinforcement learning task libraries

Recent research in task transfer and task clustering has necessitated the need for task similarity measures in reinforcement learning. Determining task similarity is necessary for selective transfer where only information from relevant tasks and portions of a task are transferred. Which task similarity measure to use is not immediately obvious. It can be shown that no single task similarity measure is uniformly superior. The optimal task similarity measure is dependent upon the task transfer method being employed. We define similarity in terms of tasks, and propose several possible task similarity measures, d_T, d_P, d_Q, and d_R which are based on the transfer time, policy overlap, Q-values, and reward structure respectively. We evaluate their performance in three separate experimental situations.