Developing Committee of Explorers for Reinforcement Learning Agents

摘要

In this paper, we describe an architecture and algorithms for reinforcement learning using a population of exploration control units. The control units, collectively known as committee of explorers, allow a reinforcement-learning agent to act while learning by choosing actions to perform through a combination of value predictions of the individual committee members. We discuss reasons for using committee of explorers and use results of an empirical study to provide insights to the performance of the committee approach.