Scholarlink

<p>Reinforcement learning has demonstrated significant advantages in dynamic control applications through its adaptive and self-learning capabilities. To address the nonlinear control challenges in valve-controlled asymmetric hydraulic cylinders, an enhanced twin delayed deep deterministic policy gradient (TD3) algorithm has been developed as a superior alternative to conventional proportional integral derivative (PID) control. The improved architecture incorporates a self-attention layer within the TD3 actor network, enabling the agent to effectively capture critical state-action relationships. Furthermore, the algorithm implements reward centering to normalize reward signals and introduces a novel target Q-value formulation to mitigate the inherent underestimation bias in TD3. Experimental results verify that the improved TD3 algorithm significantly reduces reward signal variance and enhances learning performance over the original TD3, while also achieving lower acceleration metrics. The proposed algorithm delivers control performance comparable to a conventional fixed-gain PID controller.</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

Research on control strategy of valve-controlled asymmetric cylinder based on improved TD3 algorithm

  • 期刊论文

  • 发表日期：

    2026年3月9日

查看全文
  • Yu Tian,
  • Ruiping Xiong,
  • Haoxin Gu,
  • Huizhen Wang,
  • Fulin Liao,
  • Bo Peng

摘要

Reinforcement learning has demonstrated significant advantages in dynamic control applications through its adaptive and self-learning capabilities. To address the nonlinear control challenges in valve-controlled asymmetric hydraulic cylinders, an enhanced twin delayed deep deterministic policy gradient (TD3) algorithm has been developed as a superior alternative to conventional proportional integral derivative (PID) control. The improved architecture incorporates a self-attention layer within the TD3 actor network, enabling the agent to effectively capture critical state-action relationships. Furthermore, the algorithm implements reward centering to normalize reward signals and introduces a novel target Q-value formulation to mitigate the inherent underestimation bias in TD3. Experimental results verify that the improved TD3 algorithm significantly reduces reward signal variance and enhances learning performance over the original TD3, while also achieving lower acceleration metrics. The proposed algorithm delivers control performance comparable to a conventional fixed-gain PID controller.