MarcoPolo: A Reinforcement Learning System considering tradeoff exploration and exploitation under Marcovian Environments

摘要

Reinforcement learning is a kind of machine learning. It aims to adapt an agent to a given environment with a clue to rewards. We consider that ideal reinforcement learning systems are to get some rewards even at an early learning systems are to get some rewards even at an early learning phase and to get more rewards as exploration of the environment propceeds. In this paper, we propose a unified learning system: MarcoPolo that takes account of both getting rewards by Profit Sharing or Policy Iteration and exploring the environment by k-Certainty Exploration Method. MarcoPolo can realize any tradeoff between exploitation and exploration through whole learning processes. By applying MarocPolo to numerical examples, its effectiveness is shown.