航空学报 > 2019, Vol. 40 Issue (3): 322493-322493   doi: 10.7527/S1000-6893.2018.22493

基于神经网络和人工势场的协同博弈路径规划

张菁1,2, 何友1,3, 彭应宁1, 李刚1   

  1. 1. 清华大学 电子工程系, 北京 100084;
    2. 复杂航空系统仿真重点实验室, 北京 100076;
    3. 海军航空大学 信息融合研究所, 烟台 264001
  • 收稿日期:2018-06-28 修回日期:2018-07-31 出版日期:2019-03-15 发布日期:2018-09-17
  • 通讯作者: 张菁 E-mail:jingzhang2016@tsinghua.edu.cn
  • 基金资助:
    中国博士后科学基金(2018M631483)

Neural network and artificial potential field based cooperative and adversarial path planning

ZHANG Jing1,2, HE You1,3, PENG Yingning1, LI Gang1   

  1. 1. Department of Electronic Engineering, Tsinghua University, Beijing 100084, China;
    2. Science and Technology on Complex Aviation Systems Simulation Laboratory, Beijing 100076, China;
    3. Research Institute of Information Fusion, Naval Aeronautical University, Yantai 264001, China
  • Received:2018-06-28 Revised:2018-07-31 Online:2019-03-15 Published:2018-09-17
  • Supported by:
    Postdoctoral Science Foundation of China (2018M631483)

摘要: 协同博弈路径规划是空战自主决策、机器人体育比赛等应用场景中的重要问题,其难点在于对环境对抗性反馈的实时自适应和多智能体的相互配合。提出一种基于神经网络和人工势场的协同博弈路径规划方法,使用反向传播(BP)神经网络自适应调整人工势场函数系数,并将人工势场作为神经网络输出端的特征提取。为解决真实样本质量和数量不足的问题,基于遗传算法仿真生成样本数据用于神经网络训练,并通过滚动时域的思路面向动态博弈优化样本性能。从样本数据中提炼出距离差与航向差以反映协同和博弈特性,利用神经网络的黑盒特性和学习能力解决协同博弈问题。应用于二对一反隐身超视距空战路径规划,比经典人工势场法有明显性能提升,且计算开销可接受,计算复杂度分析表明该方法可以较好扩展到多机对抗场景。

关键词: 协同博弈路径规划, 仿真生成样本, 神经网络, 人工势场, 遗传算法, 滚动时域优化, 超视距空战, 反隐身

Abstract: Cooperative and adversarial path planning is a significant issue in scenarios such as air combat and sport games. The challenge is the adaptation for dynamic feedback and the cooperation between multi-agents. A neural network and artificial potential field based method is proposed, in which the potential gain coefficient of the artificial potential field is adaptively adjusted by the Back Propagation (BP) neural network. The artificial potential field can be seen as feature extraction for the neural network on its output phase. To face the issue of insufficient natural samples for training the neural network, the sample is generated by simulation and optimized by a genetic algorithm and receding horizon optimization. The "different of distance" and "different of heading" are defined to show the characters of cooperative and adversarial path planning, and the black box feature and learning capacity of neural network are well exploited for cooperative and adversarial path planning. This method is evaluated in a two on one anti-stealth beyond-visual-range air combat, and shows significant improvement of performance and affordable costs. The computational complex analysis shows that our algorithm is scalable for multi-aircrafts cases.

Key words: cooperative and adversarial path planning, simulated sample generation, neural network, artificial potential field, genetic algorithm, receding horizon optimization, beyond-visual-range air combat, anti-stealth

中图分类号: