Abstract
Adaptive traffic signal control (ATSC) is a critical component of smart city infrastructure, aiming to reduce congestion and improve traffic flow. This study evaluates the effectiveness of reinforcement learning (RL) approaches for ATSC compared to traditional rule-based methods. Using a simulation environment based on real-world traffic data from an urban intersection, we implement and compare four RL algorithms: Q-learning, Deep Q-Network (DQN), Proximal Policy Optimization (PPO), and Multi-Agent Deep Reinforcement Learning (MADRL). Performance metrics include average waiting time, queue length, throughput, and environmental impact. Results demonstrate that RL-based methods significantly outperform rule-based systems, with DQN and PPO achieving up to 35% reduction in average waiting time and 28% improvement in throughput. MADRL further enhances performance by 12% over single-agent approaches in multi-intersection scenarios. Sensitivity analysis reveals that state representation and reward design critically influence learning efficiency. The findings highlight the potential of RL for scalable, real-time traffic management in smart cities.