Dynamic Weight-based Multi-Objective Reward Architecture for Adaptive Traffic Signal Control System

An Adaptive Traffic Signal Control (ATSC) system uses real-time traffic information to control traffic lights and makes the public transport system more reliable and accessible. Deep Reinforcement Learning (DRL) has recently demonstrated its use in resolving traffic signal control problems. However, designing a good reward function is one of the most crucial aspects of DRL since the system learns to make proper decisions based on reward. Furthermore, the multi-objective reward function is preferable for the ATSC system, which is more challenging than designing a single objective. The existing multi-objective reward functions use pre-defined fixed weights to combine the multiple parameters, which requires rigorous training and cannot represent the actual impact of the parameters. To solve this problem, we proposed a new reward architecture called Dynamic Weights Multi-objective Reward Architecture (DWMORA) for ATSC. It calculates the weights instantly based on the current traffic condition to ensure the actual impact of the parameters. A comparative result study of the proposed approach with several existing reward functions shows the improvement of the road traffic in terms of waiting time, travel time, and halting number.