모바일 게임플레이 행동 정책 학습을 위한 YOLOv3 기반 강화학습

Abstract

This paper proposes a reinforcement learning model that constructs a sequential behavioral decision policy for playing a game by extracting feature points in an environment in which a game image is given. In this paper, we propose a method of optimizing performance through state domain reduction, transfer learning, and multi-agent-based modeling to obtain the maximum score available for game environments that must continue their actions and have time limitations in decision making. These methods were implemented for the ‘Timberman’ game environment and experimented with learning performance by applying them as a player’s behavioral policy to evaluate the trained model.