Abstract

Reinforcement Learning (RL) handles policy search problems: searching a mapping from state space to action space. However RL is based on gradient methods and as such, cannot deal with problems with multimodal landscape. In contrast, though Genetic Algorithm (GA) is promising to deal with them, it seems to be unsuitable for policy search problems from the viewpoint of the cost of evaluation. Minimal Generation Gap (MGG), used as a generation-alternation model in GA, generates many offspring from two or more parents selected from a population. Therefore, evaluating policies of generated offspring requires much trial and error (i.e. interaction between an agent and an environment). In this paper, we incorporate importance sampling into the framework of MGG in order to reduce the cost of evaluation on policy search. The proposed techniques are applied to Markov Decision Process (MDP) with multimodal landscape. The experimental results show that these techniques can reduce the number of interaction between an agent and an environment, and also mean that MGG and importance sampling are good for each other.