强化学习驱动的时间约束控制研究综述:从单系统到多智能体系统OA

Reinforcement learning(RL)methods provide a new data-driven design perspective for the optimal control of nonlinear systems and the cooperative control of multi-agent systems,especially in scenarios with incomplete models,uncertain environments,and complex control objectives.Meanwhile,time-constrained control can characterize system convergence speed,transient response,and task completion deadlines,and has become an important approach for improving the time-domain performance of intelligent control systems.Focusing on RL-driven time-constrained control,this paper first introduces the basic ideas,typical structures,and main limitations of RL-based control in control applications.Then,representative time-constrained concepts,including finite-time,fixed-time,predefined-time,and prescribed-time control,are reviewed,and the necessity and implementation approaches of integrating these concepts with RL are analyzed.On this basis,from the perspectives of single systems and multi-agent systems,this paper summarizes the research progress of RL in time-constrained stabilization,optimal control,trajectory tracking,consensus control,formation coordination,and game-theoretic confrontation.The main characteristics of existing methods are also discussed in terms of system modeling,performance characterization,and control architecture.Finally,the remaining problems in theoretical analysis,constraint handling,distributed implementation,and engineering applications are summarized,and future research directions are outlined.