Structure search of probabilistic models and data correction for EDA-RL

We have proposed a novel Estimation of Distribution Algorithm for solving reinforcement learning problems: EDA-RL. The EDA-RL can perform well if the complexity of the structure of the probabilistic model is adapted to the difficulty of given problems. Therefore, this paper proposes a structure search method of the probabilistic model in the EDA-RL as in conventional EDA taking account multivariate dependencies. Moreover, a data correction method by eliminating loops of state transitions is also proposed. Computational simulations on maze problems, which have several perceptual aliasing states, show the effectiveness of the proposed method.