Adaptive traffic signal control based on reinforcement learning with consideration of vehicle and pedestrian flows
Abstract
This paper addresses the problem of adaptive traffic signal control with joint consideration of vehicle and pedestrian flows. A Deep Q-Network-based method is proposed, in which the agent decides whether to keep the current traffic signal phase or switch to the next one based on the current state of the intersection. This state includes traffic lane congestion, the number of pedestrians in the crossing zone, and the current control phase. The main contribution of the paper is the development of a multi-component adaptive reward function. It includes vehicle delays, pedestrian waiting times, intersection capacity, and a penalty for excessive phase switching. When pedestrians accumulate in the crossing area, the pedestrian-related component of the reward is increased. The experimental evaluation was carried out in the SUMO microscopic traffic simulation environment. The proposed method was compared with fixed traffic light control. The results showed a reduction in average pedestrian and vehicle wait times, shorter queue lengths, and an increase in intersection capacity. These results confirm that reinforcement learning combined with an adaptive reward function can be used for more flexible and balanced urban intersection management.