基于深度强化学习的无人战车自主行为决策  

Autonomous Behavior Decision of Unmanned Combat Vehicle Based on Deep Reinforcement Learning

作　　者：张耀 武富春[1] 王明[1] 段宏 张昭 王海龙 ZHANG Yao;WU Fu-chun;WANG Ming;DUAN Hong;ZHANG Zhao;WANG Hai-long(North Automatic Control Technology Institute,Taiyuan 030006,China)

机构地区：[1]北方自动控制技术研究所,太原030006

出　　处：《火力与指挥控制》2021年第4期72-77,共6页Fire Control & Command Control

基　　金：兵器工业联合基金资助项目(6141B011504)。

摘　　要：针对高动态强对抗战场环境下,无人战车面临的自主行为决策问题,分析了未来陆战场无人战车实际作战需求,构建了基于马尔可夫决策过程的自主行为决策模型,提出了一种深度强化学习结合行为树的方法,利用行为树的逻辑规则与先验知识降低强化学习问题的难度,保证收敛性和鲁棒性,同时使行为决策模型具有学习能力。构建典型作战场景,验证深度强化学习结合行为树的无人战车自主行为决策方法的有效性。Aiming at the problems of autonomous behavior decision-making faced by unmanned combat vehicles in a highly dynamic and strong confrontation battlefield environment,the actual combat requirements of future unmanned combat vehicles on land battlefields are analyzed,and an autonomous behavior decision-making model based on the Markov decision-making process is constructed.The method of deep reinforcement learning combined with behavior trees is proposed,the logic rules and prior knowledge of the behavior tree are used to reduce the difficulty of reinforcement learning problems and to ensure convergence and robustness,and to make behavior decision models have learning capabilities to deal with emergency situations.A typical combat scenario is constructed to verify the validity of the autonomous behavior decision-making method of unmanned combat vehicles combined with deep reinforcement learning and behavior trees.

关 键 词：无人战车 火力打击决策 强化学习 行为树 

分 类 号：TJ811[兵器科学与技术—武器系统与运用工程]