Stochastic Information Expressed in an Exponential Mixture Model of Bayesian Networks

Abstract

How to adapt agents to dynamic environments has been a crucial issue in the machine learning field. Many reinforcement learning (RL) methods have been proposed to address the above problem. The data obtained while robots acquire policies by RL are generally immediately revoked, or are eliminated shortly after a part of them is used to share the reward; however, they are considered to have potentials to reflect the characteristics of a correspondent environment. We have leveraged these data in the form of Bayesian Network (BN), and have proposed a system improving RL agents' policies with a mixture model of BNs. Our system allows agents to improve their policies by the information derived from the mixture model. This study aims to investigate the adaptability of our policy- improving system to dynamically-switched ""actual"" environments, and the relationships between the knowledge representations in the mixture and the policy in correspondent environments.