State evaluation strategy for exemplar-based policy optimization of dynamic decision problems

Direct policy search (DPS) that optimizes the parameters of a decision making model, combined with evolutionary algorithms which enable robust optimization, is a promising approach to dynamic decision problems. Exemplar- based policy (EBP) optimization is a novel framework for DPS in which the policy is composed of a set of exemplars and a case- based action selector, with the set of exemplars being refined and evolved using a GA. In this paper, state evaluation type EBP representations are proposed for the problem class whose state transition can be predicted. For example, the vector-real representation defines pairs of feature vector and its desirability as exemplars, and evaluate the predicted next states using the exemplars. The state evaluation type EBP-based optimization procedures are shown to be superior to conventional state-action type EBP optimization through application to the Tetris game.