Learning Form Experience: A Bayesian Network Based Reinforcement Learning Approach

Agent completely depends on trail-and-error to learn the optimal policy is the major reason to make reinforcement learning being slow and time consuming. Excepting for trail-and-error, human can also take advantage of prior learned experience to plan and accelerate subsequent learning. We propose an approach to model agent's learning experience by Bayesian Network, which can be used to shape agent for bias exploration towards the most promising regions of state space and thereby reduces exploration and accelerate learning. The experiment results on Grid-World problem show our approach can significantly improve agent's performance and shorten learning time. More importantly, our approach makes agent can take advantage of its learning experience to plan and accelerate learning.