Scholarlink

<p>Autonomous driving technology, as the core pillar of vehicle intelligence, shows great potential in improving road network vehicle speed, ride comfort, and traffic safety. However, this technology still faces many serious challenges, such as the high complexity of scenarios, insufficient coupling between decision-making and control, and the difficulty of ensuring safety in environments with dense vehicle interactions, especially in scenarios that require high computational power and real-time response. To address these challenges, this study proposes an integrated decision-making and control framework based on deep reinforcement learning. This framework cleverly combines the discrete double delayed deep Q-network (D3QN) for lane-changing decisions and the continuous twin delayed deep deterministic policy gradient (TD3) algorithm for car-following control. Additionally, the study introduces a context-aware lane-changing benefit function and a coordinated longitudinal control strategy to balance individual vehicle performance with the optimization of overall traffic flow. To achieve efficient coordination between the two layers, D3QN and TD3 are trained in a coupled manner. Experimental results on the simulation of urban mobility simulation platform show that under conditions of 2000 pcu/h traffic flow and 50% penetration rate of connected and autonomous vehicles (CAVs), this method can increase the average road network vehicle speed by up to 4.58% and improve the cruising speed of CAVs by 4.32% compared to the baseline model, with this advantage becoming more pronounced as traffic flow increases. These achievements fully validate that this framework can help CAVs make more intelligent and real-time decisions, thereby significantly improving the efficiency of the entire traffic system.</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

Integrated control strategy for autonomous vehicle decision-making based on deep reinforcement learning

  • 期刊论文

  • 发表日期：

    2025年8月16日

查看全文
  • Dexin Liu,
  • Xin Zhao,
  • Tenghui Ge,
  • Li Song,
  • Xuequan Zhang

摘要

Autonomous driving technology, as the core pillar of vehicle intelligence, shows great potential in improving road network vehicle speed, ride comfort, and traffic safety. However, this technology still faces many serious challenges, such as the high complexity of scenarios, insufficient coupling between decision-making and control, and the difficulty of ensuring safety in environments with dense vehicle interactions, especially in scenarios that require high computational power and real-time response. To address these challenges, this study proposes an integrated decision-making and control framework based on deep reinforcement learning. This framework cleverly combines the discrete double delayed deep Q-network (D3QN) for lane-changing decisions and the continuous twin delayed deep deterministic policy gradient (TD3) algorithm for car-following control. Additionally, the study introduces a context-aware lane-changing benefit function and a coordinated longitudinal control strategy to balance individual vehicle performance with the optimization of overall traffic flow. To achieve efficient coordination between the two layers, D3QN and TD3 are trained in a coupled manner. Experimental results on the simulation of urban mobility simulation platform show that under conditions of 2000 pcu/h traffic flow and 50% penetration rate of connected and autonomous vehicles (CAVs), this method can increase the average road network vehicle speed by up to 4.58% and improve the cruising speed of CAVs by 4.32% compared to the baseline model, with this advantage becoming more pronounced as traffic flow increases. These achievements fully validate that this framework can help CAVs make more intelligent and real-time decisions, thereby significantly improving the efficiency of the entire traffic system.