In-Depth Evaluation of Reinforcement Learning Based Adaptive Traffic Signal Control Using TSCLAB

Adaptive Traffic Signal Control (ATSC) is today widely applied for managing traffic on signalized intersections due to its capability to reduce congestion. ATSC changes the signal programs in real-time according to the measured current incoming traffic flows when ATSC is applied. This results in an improvement in the throughput of urban networks. However, prior to the implementation of such systems they be evaluated. Evaluation of the effectiveness of complex ATSC is still a challenge and presents an open problem. For the evaluation, different measures of effectiveness to gather in-depth insight into the traffic situations of the controlled signalized intersection are needed. In this paper, an augmented version of the previously developed MATLAB based tool TSCLab (Traffic Signal Control Laboratory) is applied to evaluate a newly proposed ATSC based on self-organizing maps and reinforcement learning. The performance of the mentioned ATSC is evaluated using appropriately chosen measures of effectiveness obtained in real-time using a microscopic simulation environment based on VISSIM and a realistic traffic scenario. Obtained simulation results reveal that ATSC uses shorter phase and cycle duration, achieving a lower green time utilization but also shorter queue lengths, thus improving the throughput of the analyzed intersection compared to the existing fixed-time signal control.