Reinforcement Learning with Expectation and Action Augmented States in Partially Observable Environment

抄録

The problem of developing good or optimal policies for partially observable Markov decision processes (POMDP) remains one of the most alluring areas of research in artificial intelligence. Encourage by the way how we (humans) form expectations from past experiences and how our decisions and behaviour are affected with our expectations, this paper proposes a method called expectation and action augmented states (EAAS) in reinforcement learning aimed to discover good or near optimal policies in partially observable environment. The method uses the concept of expectation to give distinction between aliased states. It works by augmenting the agent’s observation with its expectation of that observation. Two problems from the literature were used to test the proposed method. The results show promising characteristics of the method as compared to some methods currently being used in this domain.