Bayesian multi-task reinforcement learning

Abstract

In this paper, we will discuss a Bayesian learning technique in multitask reinforcement learning situations. Multitask reinforcement learning aims to make learning more efficient and faster by sharing the information that is learned in various tasks, that is, information that can be commonly applied in multiple tasks. In recent research, it has been possible to share initial policy information by restricting the Kullback-Leibler divergence between the learning policies of each task and the initial policy shared by all tasks. However, if there are environments with conflicting goals in a given set of environments, this initial policy sharing scheme may prevent agents from learning the conflicting environments because it transfers unprofitable knowledge to each task. To solve this problem, this paper suggests the Bayesian methodology in learning an initial policy and a method to learn Bayesian deep artificial neural network policies. By using the Bayesian methodology the distribution of initial policy parameter is learned. So each task-specific policy can deal with the uncertainty of the knowledge transferred from the initial policy. The proposed method is evaluated in the grid world environments and shows more stable and higher performance than the prior work.