障碍环境下基于生物捕食逃逸行为的多运动体边界防御博弈方法

基于微分对策的追逃博弈和疆土防御问题是多智能体对抗博弈控制的关键问题之一.本文研究了含障碍物的有界区域中多运动体边界防御博弈方法.首先通过对自然界中生物的捕食逃逸行为进行分析,对多运动体边界防御博弈系统进行了建模,得到不同情况下博弈终止条件和价值函数.其次,本文对含障碍环境下博弈双方的主导区域和界栅面进行了分析,并与无障碍情况进行了对比.最后,数值仿真验证了本文提出的多运动体边界防御方法满足微分对策中的鞍点策略条件和有效性.     

仿灰狼合作捕食行为的无人机集群动态任务分配

大规模无人机集群可以承担单个无人机无法完成的复杂任务, 为提高无人机集群执行任务的成功率与可 靠性, 综合考虑资源均衡与多目标分配的任务规划必不可少. 灰狼是一种典型的团队合作狩猎动物, 在捕食过程中 体现出明显的社会层级结构与合作捕食行为. 通过分析灰狼群体的狩猎过程, 对灰狼的群体交互机制与合作捕食行 为建模; 在此基础上, 将灰狼合作捕食行为机制映射至无人机集群动态任务分配中, 给出了无人机集群动态任务分 配流程. 仿真结果表明, 当无人机集群数量与任务节点关系不同时, 所提无人机集群动态任务分配方法在资源均衡 度与有效性方面均有明显优势.     

基于MASAC的无人机集群对抗博弈方法

随着无人机智能化水平的提高和集群控制技术的发展,无人机集群对抗智能决策方法将成为未来无人机作战的关键技术.无人机集群对抗学习环境具有维度高、非线性、信息有缺失、动作空间连续等复杂特点.近年来,以深度学习和强化学习为代表的人工智能技术取得了很大突破,深度强化学习在解决复杂环境下智能决策问题方面展现出了不俗能力.本文受多智能体集中式训练–分布式执行框架和最大化策略熵思想的启发,提出一种基于非完全信息的多智能体柔性行动器–评判器(multi-agent soft actor-critic, MASAC)深度强化学习方法,建立基于多智能体深度强化学习的无人机集群对抗博弈模型,构建连续空间多无人机作战环境,对红蓝双方无人机集群的非对称性对抗进行仿真实验,实验结果表明MASAC优于现有流行的多智能体深度强化学习方法,能使博弈双方收敛到收益更高的博弈均衡点.进一步对MASAC的收敛情况进行实验和分析,结果显示MASAC具有良好的收敛性和稳定性,能够保证MASAC在无人机集群对抗智能决策方面的实用性.     

基于多agent系统的大规模无人机集群对抗

本文将多agent系统引入到大规模无人机集群对抗决策系统中,给出了基于多agent系统的大规模无人机集群对抗决策方法.将机群中的每个无人机视为一个独立agent,建立了无人机运动模型,为无人机设计了独立的个体行为集,并针对每种行为给出了决策方法.通过每个个体无人机对其邻域环境的作用,涌现出宏观的集群对抗(作战)效果.使用MATLAB仿真软件对所设计的大规模无人机集群对抗方法进行了仿真,验证了所设计的基于多agent系统的大规模无人机集群对抗决策方法的有效性.     

基于多域对抗学习的无人机目标跟踪算法

针对无人机视频跟踪中正样本不足和单帧强判别特征易导致分类器过拟合的问题,提出一种基于多域对抗学习的实时无人机目标跟踪算法.将生成对抗网络引入到多域学习的特征生成中,利用对抗学习提高特征提取的鲁棒性;在卷积层中加入具有不同扩展系数的空洞卷积进行多尺度特征抽取,构建具有不同感受野的特征提取模块;在交叉熵损失函数中添加调制因子解决正负样本数量不平衡的问题.实验结果表明,该算法的跟踪精度、成功率均得到了提高.     

规则耦合下的多异构子网络MADDPG博弈对抗算法

针对多无人机博弈对抗过程中无人机数量动态衰减问题和传统深度强化学习算法中的稀疏奖励问题及无效经验抽取频率过高问题,本文以攻防能力及通信范围受限条件下的多无人机博弈对抗任务为研究背景,构建了红、蓝两方无人机群的博弈对抗模型,在多智能体深度确定性策略梯度(multi-agent deep deterministic policy gradient, MADDPG)算法的Actor-Critic框架下,根据博弈环境的特点对原始的MADDPG算法进行改进。为了进一步提升算法对有效经验的探索和利用,本文构建了规则耦合模块以在无人机的决策过程中对Actor网络进行辅助。仿真实验表明,本文设计的算法在收敛速度、学习效率和稳定性方面都取了一定的提升,异构子网络的引入使算法更适用于无人机数量动态衰减的博弈场景;奖励势函数和重要性权重耦合的优先经验回放方法提升了经验差异的细化程度及优势经验利用率;规则耦合模块的引入实现了无人机决策网络对先验知识的有效利用。     

仿鹰群智能的无人机集群协同对抗飞行验证

自主能力强且低成本的无人机集群协同对抗,是无人机集群对抗中打击敌方攻击防御体系和拦截敌方入侵机群的一个重要手段.哈里斯鹰是一种集群狩猎的猛禽,集群狩猎对于哈里斯鹰获取维持生命活动所需的能量具有重要意义.从无人机集群协同对抗任务与哈里斯鹰协同狩猎行为相似性出发,本文提出一种仿鹰群智能的无人机集群协同对抗方法.首先通过分析鹰群的集群狩猎行为,建立鹰群智能行为机制,并将其映射到无人机集群协同对抗行为中;在该模型的基础上,利用李雅普诺夫导航向量场控制无人机的运动状态,使得我方无人机能够以恒定的速度收敛到预定的轨迹上,完成对敌方无人机的对抗打击;最后,搭建无人机集群验证平台,对所设计的仿鹰群无人机集群协同对抗模型进行外场飞行验证,试验结果验证了本文所设计的模型在无人机对抗环境中的可行性与有效性.     

改进自组织映射的多无人机协同任务分配方法

针对现有算法对多无人机（UAV）协同进行多任务分配时存在负载均衡和执行效率方面的不足,提出一种改进的自组织映射（ISOM）算法。该算法根据飞行时间和任务执行时间设计了UAV的负载均衡度,以提升任务完成的效率;还设计了新的非线性变化的学习率和邻域函数保证ISOM算法的稳定性和快速收敛。然后,在不同任务环境对ISOM算法进行了有效性验证。实验结果表明,与结合遗传算法的粒子群优化（GA-PSO）、Gurobi和ORTools算法相比,ISOM算法的任务完成时间可分别减少15.5%、12.7%和7.3%;在TSPLIB数据集的实例KroA100、KroA150、KroA200上进行航迹长度减小的有效性验证时,与杂草优化（IWO）算法、改进的单亲遗传算法（IPGA）和蚁群单亲遗传算法（AC-PGA）的对比结果表明,ISOM算法在无人机数量为2、3、4、5、8时,均获得了最小的航迹长度。由此可见,ISOM算法在解决多UAV协同多任务分配问题时效果显著。     

基于粒子群算法的多无人机任务分配方法

作为多无人机系统应用的一项关键技术,任务分配是一个多维互异离散变量的优化问题.采用混合整数线性规划方法构造优化函数,并利用群智算法中的粒子群算法来求最优解,这样可以解决多无人机的任务分配问题.针对互异性要求进行必要的算法改进.数值仿真实验表明,该粒子群算法可以迅速找到优化函数的最优解,从而高效地实现多无人机的任务分配.     

On One Feature of Multicriteria Differential Games

Differential two-person zero-sum games with a vector payoff function are considered. A counterexample states that a payoff function component convolution into a linear convolution and further finding saddle point results in the interior instability of a set of such solutions. It is found that such saddle points are Geoffrion saddle points for an initial multicriteria game.     

The role of information in the cop-robber game

We investigate the role of the information available to the players on the outcome of the cops and robbers game. This game takes place on a graph and players move along the edges in turns. The cops win the game if they can move onto the robber’s vertex. In the standard formulation, it is assumed that the players can “see” each other at all times. A graph G$G$ is called cop-win if a single cop can capture the robber on G$G$. We study the effect of reducing the cop’s visibility. On the positive side, with a simple argument, we show that a cop with small or no visibility can capture the robber on any cop-win graph (even if the robber still has global visibility). On the negative side, we show that the reduction in cop’s visibility can result in an exponential increase in the capture time. Finally, we start the investigation of the variant where the visibility powers of the two players are symmetrical. We show that the cop can establish eye contact with the robber on any graph and present a sufficient condition for capture. In establishing this condition, we present a characterization of graphs on which a natural greedy pursuit strategy suffices for capturing the robber.     

Environmental effects on the coevolution of pursuit and evasion strategies

The game of tag is frequently used in the study of pursuit and evasion strategies that are discovered through competitive coevolution. The aim of coevolution is to create an arms race where opposing populations cyclically evolve in incremental improvements, driving the system towards better strategies. A coevolutionary simulation of the game of tag involving two populations of agents; pursuers and evaders, is developed to investigate the effects of a boundary and two obstacles. The evolution of strategies through Chemical Genetic Programming optimizes the mapping of genotypic strings to phenotypic trees. Four experiments were conducted, distinguished by speed differentials and environmental conditions. Designing experiments to evaluate the efficacy of emergent strategies often reveal necessary steps needed for coevolutionary progress. The experiments that excluded obstacles and boundaries provided design pointers to ensure coevolutionary progress as well as a deeper understanding of strategies that emerged when obstacles and boundaries were added. In the latter, we found that an awareness of the environment and the pursuer was not critical in an evader’s strategy to survive, instead heading to the edge of the boundary or behind an obstacle in a bid to ‘throw-off or hide from the pursuer’ or simply turn in circles was often sufficient, thereby revealing possible suboptimal strategies that were environment specific. We also observed that a condition for coevolutionary progress was that the problem complexity must be surmountable by at least one population; that is, some pursuer must be able to tag an opponent. Due to the use of amino-acid building blocks in our Chemical Genetic Program, our simulations were able to achieve significant complexity in a short period of time.

A differential pursuit/evasion game of capture between an omnidirectional agent and a differential drive robot,and their winning roles

This paper addresses a differential pursuit/evasion game. The players are an omnidirectional agent (OA) and a differential drive robot (DDR). They move in an obstacle-free environment; the DDR is faster than the OA but it can only change its motion direction up to a bounded rate. First, we analyse the scenario in which the OA has as objective to capture a DDR in minimum time and the DDR wants to retard the capture as long as possible. We present the time optimal motion primitives of the players to achieve their goals. Later, we allow the agents to change the roles, namely, the DDR is allowed to play as the pursuer and the OA is allowed to play as the evader. This later analysis allows one to establish which is the winner role for each agent, based only on the initial position of the players and their maximum speed.