基于强化学习的智能空战模型研究  

Research of intelligent air combat model based on reinforcement learning

作　　者：李佳桐 卢俊元 王光耀[2] 李建勋[1] LI Jiatong;LU Junyuan;WANG Guangyao;LI Jianxun(Shanghai Jiao Tong University,Shanghai 200240;Shenyang Aircraft Design&Research Institute,Shenyang 110031,China)

机构地区：[1]上海交通大学,上海200240 [2]沈阳飞机设计研究所,辽宁沈阳110031

出　　处：《指挥控制与仿真》2024年第4期35-43,共9页Command Control & Simulation

基　　金：重点研发计划(2020YFC1512203);上海商用飞机系统工程联合研究基金(CASEF-2022-MQ01)。

摘　　要：人工智能的进步深刻改变了各行各业,智能空战就是其中的典型代表之一。根据空战游戏的特点,构建了智能空战模型,主要包括样本集合的获取以及适用于空战决策的网络模型选择。结合空战连续状态、连续动作、现有战术动作等的特点,通过对比多个智能学习网络模型,选择DQN算法作为智能空战的模型,同时借助飞行模拟游戏DCS,采取与游戏内自带敌人对战的方式进行动态交互训练,得到能够在一定程度上操作战机作战的模型与具有参考价值的空战案例,通过对这些案例的分析,形成了胜/负/平局三类样本数据集。仿真结果表明,本文所构建的智能空战模型不仅有助于生成新的对策案例样本,而且有助于丰富空战战术。The development in artificial intelligence has dramatically changed all industries,among which AI-assisted air combat is a representative case of success.An Intelligent air combat model that consists of the attainment of samples and a decision-making model is constructed in connection with air combat simulator.Considering the characteristics of air combat continuous states and actions,DQN algorithm is selected as the model of intelligent air combat by comparison among several algorithms.Meanwhile,the AI network is trained interactively with AI enemies in the air combat simulation game DCS World,resulting in a model that is able to manipulate aircraft to a degree and many cases of air combat,by analyzing which a collection of winning,losing and dual samples is derived.The result of simulation indicates that the Intelligent air combat model has certain ability to generate strategic samples and enrich tactics in air combat environments.

关 键 词：空战 强化学习 飞行模拟游戏 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]