Traffic Light Cycle Control using Deep Reinforcement Technique

Traffic obstruction is one of the major problems faced by most of the metropolises in the dynamic road scene. Most of the roads in various cities (especially India) are poorly designed and the lack of separate footpaths are many of the reasons for the traffic congestion. Traffic lights deployed at most of the road intersections are inefficient, causing various issues including waiting time of vehicles and energy wastage. In the proposed approach, a learning model using the deep reinforcement technique to solve the issue by managing the traffic light cycle has been focused. The traffic lights collect the data from road intersections through sensor networks which is then represented as states small square shaped grids. Further, the duration changes of traffic lights are modelled through Markov decision method and then the formed states are mapped to reward through a convolutional neural network (CNN). Moreover, the behavior of the presented approach is expanded using the dueling network, the double Q-network (DQN), the target network and the prioritized experience replay. The model is evaluated through simulation on Simulation on Urban Mobility (SUMO) simulator for traffic lights cycle control.