Heterogeneous traffic intersections control design based on reinforcement learning

[摘要] Traffic light control is a cost-effective method to alleviate traffic congestion and deep reinforcement learning (DRL) that is increasingly favored as a method for real-time traffic light control. However, the complexities of modern urban intersections, including crossroads and T-junctions, pose challenges for DRL-based traffic light control systems that do not work well for such heterogeneous intersections. To address this problem, a Heterogeneous Advantage Actor-Critic (HA2C) model is proposed to control traffic lights for heterogeneous intersections. First, HA2C employs an intersection structure transformation scheme to mask intersection heterogeneity. Second, it develops a two-stage approach on top of an Advantage Actor-Critic (A2C) reinforcement learning model to learn both general and structure-specific policies, leading to more accurate decisions. The extensive simulations on both synthetic and real-world maps demonstrate that HA2C outperforms the state-of-the-art models in terms of higher throughput and faster travel time, while using a smaller model size in most scenarios.

[效力级别]  Early Access

[关键词] NETWORK