Scholarlink

<p>Autonomous vehicles (AVs) have gained attention for their safety enhancements and comfortable travel. Ongoing research targets improvements in AV technology, addressing challenges like road uncertainties, weather changes, and continuous state-actions. In this paper, we propose “Escalated DDPG,” an extension of the Deep Deterministic Policy Gradient (DDPG) algorithm, designed mainly for autonomous vehicle (AV) decision-making. Our novel approach tackles key challenges encountered with DDPG, including instability, slow convergence, and the growing complexity of AV environments. By upgrading action selection and learning policies based on consecutive actions and states, Escalated DDPG enhances convergence speed while maintaining a balanced exploration-exploitation trade-off. We conduct experiments in a gym environment, comparing the performance of our method with traditional DDPG. Results illustrate the superior accuracy and adaptability of Escalated DDPG in handling decision-making tasks involving continuous action and state spaces, even in complex scenarios. The findings in this paper contribute to advancing AV technology, enhancing their decision-making capabilities, and enabling more efficient and reliable autonomous driving systems.</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

Upgraded decision making in continuous domains for autonomous vehicles in high complexity scenarios using escalated DDPG

  • 期刊论文

  • 发表日期：

    2025年4月8日

查看全文
  • Khouloud Zouaidia,
  • Med Saber Rais,
  • Lamine Bougueroua

摘要

Autonomous vehicles (AVs) have gained attention for their safety enhancements and comfortable travel. Ongoing research targets improvements in AV technology, addressing challenges like road uncertainties, weather changes, and continuous state-actions. In this paper, we propose “Escalated DDPG,” an extension of the Deep Deterministic Policy Gradient (DDPG) algorithm, designed mainly for autonomous vehicle (AV) decision-making. Our novel approach tackles key challenges encountered with DDPG, including instability, slow convergence, and the growing complexity of AV environments. By upgrading action selection and learning policies based on consecutive actions and states, Escalated DDPG enhances convergence speed while maintaining a balanced exploration-exploitation trade-off. We conduct experiments in a gym environment, comparing the performance of our method with traditional DDPG. Results illustrate the superior accuracy and adaptability of Escalated DDPG in handling decision-making tasks involving continuous action and state spaces, even in complex scenarios. The findings in this paper contribute to advancing AV technology, enhancing their decision-making capabilities, and enabling more efficient and reliable autonomous driving systems.