摘要
当前游戏中非玩家角色(Non-player Character,NPC)的行为主要基于随机决策或者传统的预定义行为决策,该方法的NPC不具有对游戏环境的自主学习能力.本文研究的目的是探索将强化学习方法应用于提高游戏NPC智能,使NPC在游戏过程中能实时地学习和适应演进的游戏环境,产生最合适的行为策略来响应玩家.本文提出一种动态训练强化学习的探索率参数方法,并将该方法应用于经典的Bomber Man游戏中.实验结果表明,该方法训练的NPC比非强化学习和传统强化学习训练的NPC具有更高的智能.
Traditional non-player character （NPC） strategies are developed mainly based on stochastic decision or predefined behavior decision and these methods lack the capability of automatic learning. The purpose of the research is to exploring the application of reinforcement learning techniques in improving NPC intelligence, i.e. , producing the optimal NPC strategy that enables NPC to learn and adjust itself to game context. Specifically, the authors first presented a method of dynamically training exploration rate of reinforcement learning, and then applied the method into a classical game ＂Bomber＂. The results show that the presented method can obtain better NPC intelligence compared to traditional reinforcement learning methods.
出处
《四川大学学报（自然科学版）》
CAS
CSCD
北大核心
2014年第5期915-920,共6页
Journal of Sichuan University(Natural Science Edition)
基金
四川省科技支撑项目(2013GZX0138
2012GZ0091)
