An adaptive traffic signal control system based on PPO-LSTM reinforcement learning for urban intersections

Traffic congestion at signalized intersections leads to significant delays, increased fuel consumption, and elevated emissions, motivating adaptive control strategies that can respond to time-varying traffic conditions. This paper presents an adaptive traffic signal control framework based on deep reinforcement learning with recurrent temporal encoding, using a PPO agent equipped with an LSTM policy to capture short-term temporal dynamics from traffic observations. The evaluation focuses on a real-world signalized intersection case study in Ho Chi Minh City, simulated in SUMO with heterogeneous traffic composed of motorbikes, cars, and buses. To support practical deployment, a web-based monitoring architecture is developed, consisting of an API layer and an interactive dashboard for real-time visualization and logging. The proposed controller is compared with conventional fixed-time and actuated signal control baselines using operational and environmental metrics, including waiting time, queue length, throughput, fuel consumption, and CO2 emissions. The results indicate that PPO-LSTM can reduce average waiting time and queue length while improving throughput under the evaluated traffic scenario. The study highlights the practical potential of recurrent deep reinforcement learning for adaptive signal timing, while identifying multi-intersection scaling, learning-based ablations, and robustness under broader demand conditions as important directions for further evaluation.