Human Leading Behavior Learning for Multiple Autonomous Followers Under Constrained Communication Topologies
针对人类领导的多智能体系统，提出一种在线自适应逆微分博弈方法，使一个计算能力强的跟随者学习人类行为并计算其他自主跟随者的交互策略，解决通信受限下的协作控制问题。
Owing to the immaturity of current artificial intelligence techniques, practical multiagent systems (MASs) often require supervision and intervention from humans. However, it is unrealistic for a human to monitor the entire MAS and provide appropriate input in some circumstances. A viable approach is to allow a human to control an agent as the leader which in turn influences the other autonomous followers. To this end, a critical issue is how to learn human behavior to improve the autonomy of followers for collaborating with human effectively, since the autonomous followers do not have prior knowledge of human behavior. In this article, the human leading behavior learning problem is studied for a class of human-in-the-loop (HiTL) MASs that are not fully connected. A linear quadratic differential game framework is applied to formulate the collaborative control problem in the HiTL MAS where the human behavior is represented as a cost function whose weighting matrix is unknown to the followers. In the HiTL MAS, we select a follower that has strong computing power called follower 1 to learn the human behavior via an online adaptive inverse differential game (IDG) approach. Based on concurrent learning (CL) technique, an adaptive law is developed for follower 1 to determine the human feedback matrix online, and at the same time the interaction strategies for the autonomous followers are also calculated by follower 1 in case of the constrained communication topology. Subsequently, the weighting matrix in the human cost function is recovered by addressing a linear matrix inequality (LMI) optimization problem. Finally, a numerical example is presented to demonstrate the effectiveness of the proposed method.