Scholarlink

<p>Deep reinforcement learning (DRL) has emerged as a powerful approach in machine learning, widely applied to the control of nonlinear and multivariable systems such as aerial robots. However, conventional DRL algorithms still face challenges including instability in convergence, high computational complexity, and sensitivity to network configurations. In this study, a novel DRL framework is proposed based on the human activation function (HAF), inspired by neural decision-making mechanisms in the human brain. Unlike traditional continuous activation functions, HAF discretely regulates the activation of neural network layers by restricting outputs to binary values 0, 1, indicating whether a layer should be dynamically active or inactive in the information flow. The proposed architecture integrates HAF with the proximal policy optimization algorithm and is evaluated on a simulated quadrotor platform. Simulation results demonstrate that incorporating HAF significantly reduces the number of trainable parameters, enhances training stability, accelerates convergence, and improves trajectory tracking accuracy. Comparative analyses with standard DRL methods confirm the superiority of the proposed model in terms of adaptive control performance and decision-making efficiency in complex dynamic environments.</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

A novel deep reinforcement learning framework with human activation function to control quadrotor

  • 期刊论文

  • 发表日期：

    2026年1月27日

查看全文
  • Peyman Norouzi,
  • Hamed Shahbazi,
  • Keivan Torabi

摘要

Deep reinforcement learning (DRL) has emerged as a powerful approach in machine learning, widely applied to the control of nonlinear and multivariable systems such as aerial robots. However, conventional DRL algorithms still face challenges including instability in convergence, high computational complexity, and sensitivity to network configurations. In this study, a novel DRL framework is proposed based on the human activation function (HAF), inspired by neural decision-making mechanisms in the human brain. Unlike traditional continuous activation functions, HAF discretely regulates the activation of neural network layers by restricting outputs to binary values 0, 1, indicating whether a layer should be dynamically active or inactive in the information flow. The proposed architecture integrates HAF with the proximal policy optimization algorithm and is evaluated on a simulated quadrotor platform. Simulation results demonstrate that incorporating HAF significantly reduces the number of trainable parameters, enhances training stability, accelerates convergence, and improves trajectory tracking accuracy. Comparative analyses with standard DRL methods confirm the superiority of the proposed model in terms of adaptive control performance and decision-making efficiency in complex dynamic environments.