Decentralized Multi-Agent Reinforcement Learning for Adaptive Traffic Congestion Control: Integrating Deep Q-Networks with Kalman Filter-Based Prediction

Urban traffic congestion poses a significant challenge to modern cities, impacting mobility, air quality, and overall quality of life. Traditional traffic signal control systems, often based on fixed schedules or simple heuristics, struggle to adapt to dynamic traffic patterns, leading to inefficient traffic flow. This study addresses the pressing need for adaptive and efficient traffic signal control systems capable of responding to real-time traffic conditions across multiple intersections. We propose a novel multi-agent reinforcement learning approach for traffic congestion control, utilizing Deep Q-Network (DQN) algorithms integrated with Kalman filter-based traffic prediction. Our method incorporates a decentralized control architecture, a comprehensive state representation including vehicle counts and waiting times, and an expanded action space covering various traffic signal configurations. A key innovation is our reward model, which balances both congestion reduction and fairness in traffic flow across different directions.