样本级实时空中格斗决策可解释模型研究  

Interpretable Sample Level Real-time Air Combat Decision Model

作　　者：胡易航 裘旭益[2] 张彦[2] 袁伟伟[1] HU Yi-hang;QIU Xu-yi;ZHANG Yan;YUAN Wei-wei(College of Computer Science and Technology,Nanjing University of Aeronautics and Astronautics,Nanjing 211106,China;Science and Technology on Avionics Integration Laboratory,China National Aeronautical Radio Electronics Research Institute,Shanghai 200233,China)

机构地区：[1]南京航空航天大学计算机科学与技术学院,南京211106 [2]中国航空无线电电子研究所航空电子系统综合技术重点实验室,上海200233

出　　处：《小型微型计算机系统》2023年第11期2591-2596,共6页Journal of Chinese Computer Systems

基　　金：横向项目资助。

摘　　要：空中格斗决策作为空战领域重要组成部分,格斗决策的实时性、准确性直接关系到空战的成败,构建一个有效的、高速的、可解释的空中格斗模型显得十分重要.实际场景中,基于深度强化学习的空中格斗决策模型为了实现强大的学习能力,往往得到的模型参数过多,这导致了决策时间的延迟,不能较好的满足实时性强的空中格斗场景.同时深度强化学习模型的黑盒特性,使得模型在智能决策过程中的决策原因不可知,可是每一次空战决策过程中的不可解释性在空战场景中存在极大的未知风险.为了解决这些问题,本文提出一种基于高效的轻量级决策树的模型用于空战中战斗决策,该轻量级模型不仅满足可解释性同时满足空战场景中的高速决策.为了解决智能决策模型在每一次决策过程中的黑箱特性,本文采用一种局部代理技术完成空战格斗中每一次决策过程中的样本级解释.实践表明本文所构建的空中格斗模型具有较高的决策精度、速度与实战性能.Air combat decision-making is an important part in the field of air combat.The real-time and accuracy of combat decisionmaking are directly related to the success or failure of air combat.It is very important to build an effective,high-speed and interpretable air combat model.In the actual scene,the air combat decision-making model based on deep reinforcement learning often obtains too many model parameters in order to achieve strong learning ability,which leads to the delay of decision-making time and can not better meet the air combat scene with strong real-time.At the same time,the black box characteristic of the deep reinforcement learning model makes the decision-making reason of the model in the intelligent decision-making process unknown,but the unexplainability of each air combat decision-making process has a great unknown risk in the air combat scene.In order to solve these problems,we propose a model based on efficient lightweight decision tree for combat decision-making in air combat.The lightweight model not only meets the interpretability,but also meets the high-speed decision-making in air combat scene.In order to solve the black box characteristics of intelligent decision model in each decision-making process,a local agent technology is used to complete the sample level interpretation in each decision-making process in air combat combat combat.Practice shows that our air combat model has high decision-making accuracy,speed and practical performance.

关 键 词：空中格斗决策 黑盒问题 样本级解释 LIME 特征增强 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]