Adaptive traffic signal control using deep reinforcement learning: A multi-objective approach for single and multi-intersection scenarios

Traffic congestion in urban networks necessitates adaptive signal control systems that balance efficiency, safety, and environmental impact. While Deep Reinforcement Learning (DRL) has shown promise for traffic signal control (TSC), existing approaches often optimize single objectives or lack scalability for multi-intersection corridors. This study proposes a multi-objective DRL framework that simultaneously minimizes vehicle delays, traffic conflicts (using Time-to-Collision metrics), and CO₂ emissions. Six reward functions were designed and tested with three DRL algorithms (PPO, A2C, DQN) in synthetic and real-world intersections simulated in SUMO. Key results demonstrate that: (1) PPO outperformed A2C and DQN, achieving up to 21 % fewer conflicts and 7 % lower delays in high-traffic scenarios; (2) A safety-focused reward (Reward 3) reduced conflicts by 4–20 % compared to fixed-time controls in high-traffic scenarios, while multi-objective rewards (Rewards 5–6) balanced all targets effectively; and (3) Decentralized control for multi-intersection corridor reduced delays, conflicts, and emissions by 26.4 %, 26.9 %, and 12 %, respectively, surpassing centralized approaches. Real-world validation in Hangzhou, China, and Cologne, Germany, confirmed robustness, with 10–15 % delay reductions and 2.2–6.4 % lower emissions versus prior DRL models. This work advances adaptive TSC by integrating safety and sustainability into DRL optimization, offering scalable solutions for heterogeneous traffic networks.