Scholarlink

<p>This paper investigates a finite-time optimal bipartite containment control problem for multi-agent systems (MASs) with input saturation. Firstly, a command-filtered technique is applied to filter the virtual control signals to avoid the problem of “explosion of complexity”. Then, the filter and saturation losses are compensated simultaneously by skillfully constructing auxiliary systems, whose signals converge in finite time. Due to the strong nonlinearity of the Hamilton-Jacobi-Bellman equations and system dynamics, the modified identifier-actor-critic reinforcement learning (IAC-RL) algorithm is employed to approximate the unknown functions and train the optimal controller. Specifically, the cost function in the traditional IAC-RL algorithm is modified to ensure its convergence over a long time. With the help of a correction term, the updating laws of the IAC-RL neural networks are also improved to avoid premature termination during training optimal controllers. Finally, the MASs are proved to be semiglobally practically finite-time stable. The effectiveness of the proposed protocol is proved through numerical and practical examples.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Command-Filter-Based Finite-Time Bipartite Containment Control for MASs via Reinforcement Learning

  • 期刊论文

  • 发表日期：

    2026年4月15日

查看全文
  • Xinyu Qiu,
  • Zhenyou Wang,
  • Ao Luo,
  • Hui Ma,
  • Shengbing Xu

摘要

This paper investigates a finite-time optimal bipartite containment control problem for multi-agent systems (MASs) with input saturation. Firstly, a command-filtered technique is applied to filter the virtual control signals to avoid the problem of “explosion of complexity”. Then, the filter and saturation losses are compensated simultaneously by skillfully constructing auxiliary systems, whose signals converge in finite time. Due to the strong nonlinearity of the Hamilton-Jacobi-Bellman equations and system dynamics, the modified identifier-actor-critic reinforcement learning (IAC-RL) algorithm is employed to approximate the unknown functions and train the optimal controller. Specifically, the cost function in the traditional IAC-RL algorithm is modified to ensure its convergence over a long time. With the help of a correction term, the updating laws of the IAC-RL neural networks are also improved to avoid premature termination during training optimal controllers. Finally, the MASs are proved to be semiglobally practically finite-time stable. The effectiveness of the proposed protocol is proved through numerical and practical examples.