An Agent-Based Traffic Signal Control Using Reinforcement Learning Algorithm

Abstract

Traffic light control has been a significant test in most major roads in Nigeria. The control of traffic has been so poor in certain spots in Nigeria to such an extent that more timing is being distributed to zones with lesser vehicles while little timing is being allotted to zones of more vehicles. This paper presents an Agent-based system to determine the control of traffic light signals using Reinforcement Learning algorithm by applying Deep Q Learning Techniques. The Reinforcement learning algorithm was trained using a Deep Q-learning technique with a total of 4 input layers, a batch size of 100, learning rate of 0.001 and a training epoch of 800 and a gamma of 0.97. The learning environment was made up with a maximum number of steps of 5400, total numbers of car generated to be 1000, green light duration in 10, yellow light duration to be 4. The number of actions taken by the agent equals 4 on 80 different states. The system helps in reducing traffic congestion by adapting to the learning environment, therefore knowing lanes with more vehicles during and without rush hours. By this, system optimizes the green time effectively by allocating more time to lane with more vehicles during and with rush hours, therefore, reducing the average cumulative delays and average cumulative queued length of vehicles. The result showed that system is efficient in traffic signal control with an average queued vehicle length of 5 to 20 vehicle.