Traffic Signal Timing in Intersection Traffic Control Using Reinforcement Learning and Neural Networks

One of the main challenges in intersection traffic control is optimizing the timing of traffic lights to reduce vehicle stoppage and prevent severe traffic interference.Reinforcement learning algorithms are employed for optimal decision-making in dynamic conditions, while neural networks are utilized to identify complex traffic patterns. Vehicle speed and count are identified as the main parameters affecting traffic flow. The study incorporates three types of neural networks:multilayer perceptron networks, neuro-fuzzy networks, and radial basis function networks. Real data from intersections on Shahid Ashrafi Esfahani Highway were processed for modeling. The input parameters include maximum speed, average speed, speed variance, and vehicle flow volume, with traffic flow state as the main output parameter.Neural network models were implemented, trained, and evaluated using MATLAB Neural Network Toolbox (version 9.5, 2019a). For deep learning modeling, three types of deep neural networks were used: multilayer perceptron, neuro-fuzzy, and radial basis function networks, classified as deep networks. Each model was developed with two architectures and hidden layer configurations, with performance comparison based on various evaluation metrics. Correlation coefficients, absolute and relative error rates, were computed for evaluation.After modeling and analysis, the MLP2 network with two hidden layers of 16 neurons demonstrated the best performance among the three models.However, in the training dataset, the MLP1 network with two hidden layers of 20 neurons performed better. Nevertheless, in other error metrics, MLP2 showed lower values, making it the most successful model for predicting traffic congestion.