Traffic light dispersion control based on deep reinforcement learning

Abstract

The current traffic light controls are ineffective and causes a handful of problems such as congestion and pollution. This study investigates the application of deep reinforcement learning on traffic control systems to minimize congestion at traffic intersection. The traffic data from Pulai Perdana, Skudai, Johor Intersection was extracted, analysed and simulated based on the Poisson Distribution, using a simulator, Simulation of Urban Mobility (SUMO). In this research, we proposed a deep reinforcement learning model, which combines the capabilities of convolutional neural networks and reinforcement learning to control the traffic lights to increase the effectiveness of the traffic control system. The paper explains the method we used to quantify the traffic scenario into different matrices which fed to the model as states which reduces the load of computing as compared to images. After 2000 iterations of training, our deep reinforcement learning model was able to reduce the cumulative waiting time of all the vehicles at the Pulai Perdana intersection by 47.31% as compared to a fixed time algorithm and can perform even when the traffic is skewed in a different direction. When the traffic is scaled down to 50% and 20 %, the agent continues to improve the waiting time by 69.5% and 68.36 % respectively. It is proven in the experiment that a deep reinforcement learning model was able to reduce the cumulative waiting time at Pulai Perdana by 47.31%.