Smart Traffic Light Control via Reinforcement Learning

Abstract

Nowadays, the problem of traffic congestion is faced by a large number of cities, especially those metropolitan cities, due to the rapid growth of population and vehicles. Therefore, it is important to find an efficient method for optimizing the traffic light control at intersections so as to increase traffic throughput and reduce congestion. Recent trends have been toward applying reinforcement learning (RL) to traffic light control. RL allows an agent to learn directly from the environment by trial and error, compared with tradition control methods that depend heavily on prior knowledge of the traffic pattern. In this project, I used Q-learning to configure the control of traffic lights. Q-learning is a simple but useful RL algorithm and it finds the optimal policy in terms of maximizing the reward of an action in a given state. The traffic model is simulated using Anylogic which is a multimethod simulation modelling tool and provides an agent-based simulation approach. In order to make a comparison, I also implemented a fixed-time controller as the benchmark. Finally, the simulation results show that the Q-learning controller has an overall better performance than fixed-time controller.