Principled Methods for Biasing Reinforcement Learning Agents

摘要

Reinforcement learning (RL) is a powerful technique for learning in domains where there is no instructive feedback but only evaluative feedback and is rapidly expanding in industrial and research fields. One of the main limitations of RL is the slowness in convergence. Thus, several methods have been proposed to speed up RL. They involve the incorporation of prior knowledge or bias into RL. In this paper, we present a new method for incorporating bias into RL. This method extends the choosing initial Q-values method proposed by Hailu G. and Sommer G. and one kind of learning mechanism is introduced into agent. This allows for much more specific information to guide the agent which action to choose and meanwhile it is helpful to reduce the state research space. So it improves the learning performance and speed up the convergence of the learning process greatly.