Optimization of Transport Traffic in a Network of Medium Complexity Using Deep Learning With Reinforcement

Traffic flow optimization in urban settings continues to present a substantial challenge despite extensive research efforts. Current solutions typically utilize localized agents to manage traffic at individual intersections, but modern methodologies are increasingly incorporating deep learning techniques — notably the Deep Deterministic Policy Gradient (DDPG) algorithm — to effectively process complex traffic data. DDPG exhibited more stable and accelerated learning compared to Q-learning, achieving average rewards between 2 and 3.5 points, whereas Q-learning scores ranged from 1 to 2 points. The reward variance for DDPG was consistently within the range of 0.02 to 0.08, while Q-learning's variance fluctuated between 0.02 and 0.15. In the final steps, DDPG stabilized within the range of 0.04 to 0.06, while Q-learning continued to show instability, ranging from 0.03 to 0.12. The random approach significantly underperformed, further underscoring the advantages of reinforcement learning. These findings highlight DDPG's promising potential for effective traffic light control in dynamic urban environments.