E-DBRL: efficient double broad reinforcement learning for adaptive traffic signal control

Efficient traffic signal management is crucial for regulating traffic flow and fostering sustainable development within road transportation systems. To address the challenges in traffic management, numerous studies have applied the Adaptive Traffic Signal Control (ATSC) technology, using Deep Reinforcement Learning (DRL) to decrease vehicles’ average waiting times. Nonetheless, the intricate nature of DRL, characterized by its extensive parameter connections, often complicates the assurance of real-time responsiveness. Additionally, by prioritizing reduced waiting times, these methods may overlook potential rises in queue lengths, risking congestion. In this paper, we propose an Efficient Double Broad Reinforcement Learning (E-DBRL) algorithm based on a Double Broad Q-Network (Double BQN) to alleviate the overestimation of action values common in Broad Reinforcement Learning (BRL). To enhance the Quality of Experience (QoE) of drivers, we develop a new reward function that optimizes the average waiting time and the range between the longest and shortest waiting times, thus avoiding the need for dimension normalization. Moreover, we conduct simulation experiments using actual traffic data collected from Hangzhou, China. The experimental results indicate that, compared to the traditional Double DQN, the proposed E-DBRL algorithm achieves a 45.78% reduction in the average training time per round and a 5.57% increase in the average rewards.