A Reinforcement Learning Method for Traffic Signal Control at an Isolated Intersection with Pedestrian Flows

In the paper, we propose a model-based reinforcement learning algorithm, i.e., approximate dynamic programming for signal control at isolated intersections with mixed traffic: vehicles and pedestrians. The integrated optimization problem is formulated by the discrete-time dynamic process. The system state is represented by the combination of weighted vehicle queue lengths and the number of waiting pedestrians. The system action is generated in each decision step by the proposed algorithm. To solve the computation issue in conventional dynamic programming, the proposed algorithm adopts a linear approximation function that helps to quickly obtain a near-optimal solution. In simulation, we extract traffic information from the traffic simulator SUMO. The on-line traffic information is provided for the algorithm to make a signal decision. After testing various scenarios, results show that the proposed algorithm has potential control performance. We also reveal the delay changes with different weights assigned to the vehicle and pedestrian components.