Scholarlink

<p>This paper proposes a novel neural network (NN) learning control scheme based on lifelong learning in a multi-task environment for multi-input multi-output nonlinear systems. First, radial basis function neural networks are employed to approximate unknown system dynamics of different tracking tasks based on neuron dynamic-growing learning theory, which can automatically generate neurons around the NN inputs and is specially suitable for multi-tasking dynamic learning. Subsequently, the new weight updating law is proposed by design weight feedback terms, which allows the system to retrieve stored memories online when encountering the same task for better performance. The S-shaped screening function is embedded into the weight feedback terms to map larger NN weights (important) to higher feedback coefficients and smaller weights to near-zero coefficients, retaining crucial information. Compared to traditional adaptive NN control methods, the proposed weight feedback mechanism ensures that weights related to the current task are preserved online after the closed-loop system stabilization, preventing catastrophic forgetting problem when switching between tasks. Based on system stability theory, the designed control scheme can guarantee the boundedness of all closed-loop signals, the exponential convergence of NN weights, the online knowledge storage and reuse of multi-unknown dynamics in multi-task environment. The effectiveness of the suggested scheme is validated through simulation experiments on a closed two-link manipulator.</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

Lifelong neural learning control for MIMO nonlinear systems in multi-task environment

  • 期刊论文

  • 发表日期：

    2026年3月17日

查看全文
  • Mingyu Wang,
  • Min Wang

摘要

This paper proposes a novel neural network (NN) learning control scheme based on lifelong learning in a multi-task environment for multi-input multi-output nonlinear systems. First, radial basis function neural networks are employed to approximate unknown system dynamics of different tracking tasks based on neuron dynamic-growing learning theory, which can automatically generate neurons around the NN inputs and is specially suitable for multi-tasking dynamic learning. Subsequently, the new weight updating law is proposed by design weight feedback terms, which allows the system to retrieve stored memories online when encountering the same task for better performance. The S-shaped screening function is embedded into the weight feedback terms to map larger NN weights (important) to higher feedback coefficients and smaller weights to near-zero coefficients, retaining crucial information. Compared to traditional adaptive NN control methods, the proposed weight feedback mechanism ensures that weights related to the current task are preserved online after the closed-loop system stabilization, preventing catastrophic forgetting problem when switching between tasks. Based on system stability theory, the designed control scheme can guarantee the boundedness of all closed-loop signals, the exponential convergence of NN weights, the online knowledge storage and reuse of multi-unknown dynamics in multi-task environment. The effectiveness of the suggested scheme is validated through simulation experiments on a closed two-link manipulator.