摘要
探索深度强化学习在对抗作战策略上的应用,针对多智能体深度确定性策略梯度算法的局部可观测、训练较难收敛和稳定性较差的问题,分别引入长短时记忆神经网络、基于损失的优先级经验和策略梯度权重3种方法解决算法中对应问题,结合对抗作战决策场景与改进后的算法,设计3种决策实验场景。将算法与MADDPG、DDPG算法在多智能体模拟对抗实验环境中进行对比,结果表明算法在对抗决策的稳定性和效率上均有提升。
The application of deep reinforcement learning in the confrontation combat strategies is explored,aiming at the problems of partial observability,hard convergence in training,and poor stability of the Multi-Agent Deep Deterministic Policy Gradient(MADDPG)algorithm,three kinds of methods including Long Short-Term Memory(LSTM)network,priority experience replay based on loss,and policy gradient weight are introduced into the MADDPG algorithm to solve the corresponding problems,three kinds of decision-making experiment scenarios are designed with the combination of confrontation combat decision-making scenarios and the improved algorithm,and the proposed algorithm is compared with the MADDPG and DDPG algorithm in multi-agent simulation confrontation experiment environment.The results show that the proposed algorithm has improved the stability and efficiency of confrontation decision-making.
作者
刘鹏
赵建新
张宏映
高腾飞
闫涛
LIU Peng;ZHAO Jianxin;ZHANG Hongying;GAO Tengfei;YAN Tao(North Automatic Control Technology Institute,Taiyuan 030006,China)
出处
《火力与指挥控制》
CSCD
北大核心
2023年第3期132-138,145,共8页
Fire Control & Command Control
关键词
深度强化学习
对抗决策
长短时记忆网络
经验优先抽取
策略梯度
deep reinforcement learning
confrontation decision-making
long and shortterm memory networks
experience priority extraction
policy gradient