基于强化学习的三维游戏控制算法  

Control Algorithm of Three-Dimensional Game Based on Reinforcement Learning

作　　者：孟琭[1] 沈凝 祁殷俏 张昊园 MENG Lu;SHEN Ning;QI Yin-qiao;ZHANG Hao-yuan(School of Information Science&Engineering,Northeastern University,Shenyang 110819,China)

机构地区：[1]东北大学信息科学与工程学院,辽宁沈阳110819

出　　处：《东北大学学报（自然科学版）》2021年第4期478-482,493,共6页Journal of Northeastern University(Natural Science)

基　　金：国家重点研发计划项目(2018YFB2003502);国家自然科学基金资助项目(62073061);中央高校基本科研业务费专项资金资助项目(N2004020).

摘　　要：基于强化学习,设计了一个面向三维第一人称射击游戏(DOOM)的智能体,该智能体可在游戏环境下移动、射击敌人、收集物品等.本文算法结合深度学习的目标识别算法Faster RCNN与Deep Q-Networks(DQN)算法,可将DQN算法的搜索空间大大减小,从而极大提升本文算法的训练效率.在虚拟游戏平台(ViZDoom)的两个场景下(Defend_the_center和Health_gathering)进行实验,将本文算法与最新的三维射击游戏智能体算法进行比较,结果表明本文算法可以用更少的迭代次数实现更优的训练结果.Based on reinforcement learning,an agent for three-dimensional first person shooting game(DOOM)was designed.The agent can move,shoot enemies and collect objects in the game environment.The proposed algorithm combines the Faster RCNN algorithm of deep learning and the Deep Q-Networks(DQN)algorithm of reinforcement learning,which can greatly reduce the search space of DQN algorithm and improve the training efficiency of the proposed algorithm.The experiments were carried out in two scenes(Defend_the_center and Health_gathering)of the virtual game platform(ViZDoom),and the proposed algorithm was compared with the state-of-the-art three-dimensional shooting game agent algorithm.The results show that the proposed algorithm can achieve better training results with fewer iterations.

关 键 词：强化学习 深度学习 目标识别 Faster RCNN DQN 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]