REINFORCEMENT LEARNING-BASED TRAFFIC SIGNAL OPTIMIZATION FOR SMART CITIES

Abstract

Modern cities are not providing enough infrastructure, which has further aggravated congestion, commute, and environmental pollution. The traditional traffic signal systems (both fixed-time and actuated controls) lack the ability to adjust to real-time changes, and thus they have fewer choices available to them to deal with traffic. Recent developments in artificial intelligence, especially reinforcement learning (RL), yield new possibilities of adaptive data-driven traffic control. In this paper, the author examines how deep reinforcement learning (DRL) frameworks can be used to optimize traffic lights in smart cities. RL-based controllers are compared to SUMO-based simulations of fixed-time, actuated and SURTRAC-like scheduling systems at single intersections, corridors and grid networks. Results show that RL reduces delays by up to 45 per cent., reduces queues by over 40 meters, enhances throughput by 28 per cent., and reduces CO2 emissions by 19 per cent when compared with the baseline method. Further, when there is incidence, RL is stable and returns flow within 5-10 minutes, which is regarded as superior to conventional systems. The results highlight the scalability, sustainability, and resilience of traffic management provided by RL. The paper is concluded by advising on hybrid deployment approaches, their relation to connected vehicle data, and further research on equity, interpretability, and actual pilot implementations to build intelligent transportation in urban areas.

Keywords reinforcement learning, traffic signal control, deep reinforcement learning, SUMO, smart cities, sustainable mobility