EDA-RL: EDA with Conditional Random Fields for Solving Reinforcement Learning Problems

Abstract

This chapter introduces a novel Estimation of Distribution Algorithm for solving Reinforcement Learning Problems, i.e., EDA-RL. As the probabilistic model of the EDA-RL, the Conditional Random Fields proposed by Lafferty et al. are employed. The Conditional Random Fields can estimate conditional probability distributions by using Markov Network. Moreover, the structural search of probabilistic model by using X 2-test, and data correction method are examined. One of the primary features of the EDA-RL is the direct estimation of reinforcement learning agents’ policies by using the Conditional Random Fields. Another feature is that a kind of undirected graphical probabilistic model is used in the proposed method. The experimental results on Probabilistic Transition Problems and Maze Problems show the effectiveness of the EDA-RL.