Skill Learning with Empowerment in Reinforcement Learning

Abstract

This paper considers a skill learning method that does not use information about a goal set for an agent. In reinforcement learning, such methods are called intrinsic motivation methods. The proposed algorithm is based on the calculation of the mutual information between the actions of the agent and its states. This approach makes it possible to learn skills that lead to the most uniform reachability of possible states. The skills formed by the algorithm are illustrated by an example of continuous control of the agent’s movement. The usefulness of these skills in solving reinforcement learning problems is analyzed.