에피소드 매개변수 최적화를 이용한 확률게임에서의 추적정책 성능 향상

Abstract

In this paper, we introduce an optimization method to improve pursuit performance of a pursuer in a pursuit-evasion game (PEG). Pursuers build a probability map and employ a hybrid pursuit policy which combines the merits of local-max and global-max pursuit policies to search and capture evaders as soon as possible in a 2-dimensional space. We propose an episodic parameter optimization (EPO) algorithm to learn good values for the weighting parameters of a hybrid pursuit policy. The EPO algorithm is performed while many episodes of the PEG are run repeatedly and the reward of each episode is accumulated using reinforcement learning, and the candidate weighting parameter is selected in a way that maximizes the total averaged reward by using the golden section search method. We found the best pursuit policy in various situations which are the different number of evaders and the different size of spaces and analyzed results.