基于静态博弈和遗传算法的多智能体博弈策略生成方法  

Multi-agent Game Strategies Generation Method Based on Static Games and Genetic Algorithms

作　　者：刘东辉[1,2] 郑赢营 畅鑫 李艳斌[3] LIU Donghui;ZHENG Yingying;CHANG Xin;LI Yanbin(School of Management,Shijiazhuang Tiedao University,Shijiazhuang 050043,China;Research Institute of Engineering Management,Shijiazhuang Tiedao University,Shijiazhuang 050043,China;The 54th Research Institute of CETC,Shijiazhuang 050081,China)

机构地区：[1]石家庄铁道大学管理学院,河北石家庄050043 [2]石家庄铁道大学工程建设管理研究中心,河北石家庄050043 [3]中国电子科技集团公司第五十四研究所,河北石家庄050081

出　　处：《无线电工程》2024年第6期1355-1360,共6页Radio Engineering

基　　金：国家自然科学基金(71991485,71991481,71991480);中国博士后科学基金(2021 M693002)。

摘　　要：在多智能体协同对抗策略生成的过程中,奖励稀疏和神经网络参数多易导致策略生成速度慢。针对特定场景如何快速产生对抗策略这一问题,提出了一种基于静态博弈和遗传算法的多智能体博弈策略生成方法。基于静态博弈理念,对马尔科夫决策过程演化,将策略映射为一串动作组成,简化策略映射原理;对策略优化问题数学建模。以对抗结果作为目标函数,基于动作集合优化目标函数,通过优化的方法能够获得对抗结果最优的策略;给出策略优化框架,并改进遗传算法实现对于多智能体博弈策略的快速并行寻优。实验结果表明,相比于经典多智能体强化学习方法,所提方法能够高效产生多智能体博弈策略。In the process of generating multi-agent collaborative confrontation strategies,sparse rewards and numerous neural network parameters often lead to slow strategy generation.To rapidly generate confrontation strategies for specific scenarios,a method for multi-agent game strategies generation based on static games and genetic algorithms is proposed.Leveraging the concept of static games,the evolution of the Markov decision process maps the strategies to a sequence of actions,simplifying the principle of strategy mapping.Subsequently,mathematical modeling is applied to the strategy optimization problem.Using the confrontation result as the objective function and optimizing it based on the action set,the method can acquire strategies for optimal confrontation results through optimization.Then,a strategy optimization framework is presented,and genetic algorithms are improved to achieve rapid parallel optimization for multi-agent game strategies.Experimental results demonstrate that,compared to classical multi-agent reinforcement learning methods,the proposed method efficiently generates strategies for multi-agent games.

关 键 词：静态博弈 遗传算法 策略生成 

分 类 号：TN929.5[电子电信—通信与信息系统]