Scholarlink

<p>Research on cooperative control of autonomous vehicles using reinforcement learning has gained attention in the recent years. Conventional intersection control using traffic signals faces challenges in flexibly responding to changing traffic conditions, making balancing safety improvement and traffic flow optimization difficult. This study proposes an intersection agent that dynamically adjusts the entry speed of vehicles at unsignalized intersections and explores a method to acquire control policies through reinforcement learning. In the proposed method, the intersection agent observes the number of vehicles from each entry direction and implements continuous speed limits for vehicles in each direction, enabling control that responds to congestion conditions and collision risks. The Proximal Policy Optimization (PPO) deep reinforcement learning algorithm was used to learn the control policy. Simulation evaluation results showed that compared with Matsuda et al.'s method (Kishi et al. in 27:513–520, 2022), the average number of vehicles passing through was 496.4 for the proposed method and 488.6 for the conventional method, confirming that control efficiency was maintained. In addition, the number of collisions within the intersection was 9.8 on average for the proposed method, compared to 31.3 for the conventional method, achieving a significant reduction. In a four-way intersection scenario, the proposed method achieved an average vehicle throughput of 218.6 and recorded 3.3 collisions, while the baseline method yielded 181.3 and 11.0, respectively. This represents a 20.6% increase in throughput and a 70% reduction in collisions, demonstrating a significant improvement in both traffic efficiency and safety. The proposed method learned an adaptive control policy that does not impose speed limits when traffic conditions are stable, but temporarily strengthens restrictions only during congestion or high-collision risk situations, showing potential for simultaneously optimizing traffic flow and improving safety.</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

Reducing collision counts at intersections using reinforcement learning-based adaptive speed control

  • 期刊论文

  • 发表日期：

    2025年10月16日

查看全文
  • Kaito Kumagae,
  • Tenta Suzuki,
  • Mao Tobisawa,
  • Tomohiro Harada,
  • Johei Matsuoka,
  • Yuki Itoh,
  • Clive Jancen Kawaoto,
  • Kiyohiko Hattori

摘要

Research on cooperative control of autonomous vehicles using reinforcement learning has gained attention in the recent years. Conventional intersection control using traffic signals faces challenges in flexibly responding to changing traffic conditions, making balancing safety improvement and traffic flow optimization difficult. This study proposes an intersection agent that dynamically adjusts the entry speed of vehicles at unsignalized intersections and explores a method to acquire control policies through reinforcement learning. In the proposed method, the intersection agent observes the number of vehicles from each entry direction and implements continuous speed limits for vehicles in each direction, enabling control that responds to congestion conditions and collision risks. The Proximal Policy Optimization (PPO) deep reinforcement learning algorithm was used to learn the control policy. Simulation evaluation results showed that compared with Matsuda et al.'s method (Kishi et al. in 27:513–520, 2022), the average number of vehicles passing through was 496.4 for the proposed method and 488.6 for the conventional method, confirming that control efficiency was maintained. In addition, the number of collisions within the intersection was 9.8 on average for the proposed method, compared to 31.3 for the conventional method, achieving a significant reduction. In a four-way intersection scenario, the proposed method achieved an average vehicle throughput of 218.6 and recorded 3.3 collisions, while the baseline method yielded 181.3 and 11.0, respectively. This represents a 20.6% increase in throughput and a 70% reduction in collisions, demonstrating a significant improvement in both traffic efficiency and safety. The proposed method learned an adaptive control policy that does not impose speed limits when traffic conditions are stable, but temporarily strengthens restrictions only during congestion or high-collision risk situations, showing potential for simultaneously optimizing traffic flow and improving safety.