Hybrid CNN-LSTM and Proximal Policy Optimization Model for Traffic Light Control in a Multi-Agent Environment

Conventional traffic light control systems often exhibit rigid timing patterns, limited flexibility, and insufficient adaptability to changing traffic conditions. This paper addresses urban traffic management challenges, addressing researchers and professionals in traffic engineering and intelligent transportation systems. To overcome these challenges, this paper presents an innovative traffic light control method, integrating a CNN-LSTM model for traffic state prediction, combined with a Proximal Policy Optimization (PPO) algorithm for traffic light control decision-making. The model is based on a representation of intersection states through key indicators (such as active green phase, congestion levels, congestion variations, and vehicle speeds) and employs deep reinforcement learning to optimize traffic light control strategies. The adopted method is compared with fixed traffic light control approaches and a reinforcement learning (Q-learning) approach in a simulated environment using SUMO. The simulations consider diverse traffic scenarios and realistic urban conditions to ensure robust evaluation. The experimental results indicate that traffic efficiency is significantly improved by up to 92% in scenarios managing medium traffic demand, with up to 2000 vehicles per hour in the North-South scenario, while congestion indicators are substantially reduced. These improvements are achieved under conditions where traffic remains below the saturation threshold, ensuring stable flow management.The proposed method demonstrates its effectiveness in optimizing signalized intersections, significantly enhancing traffic flow in pre-saturation conditions, while opening perspectives for further research in oversaturated networks. These results illustrate the power of integrating spatiotemporal prediction and PPO-based control for dynamic and adaptive traffic management in urban networks.