Deep Reinforcement Learning-based Traffic Signal Control

Traffic congestion has become an increasingly concerning problem in modern society. Recent research has proven that Reinforcement Learning (RL) applied to Traffic Signal Control (TSC) is useful in mitigating congestion. In this paper, a model of real-world intersection with real traffic data collected in Hangzhou, China is simulated with different RL based traffic signal controllers. Two model free reinforcement learning methods are proposed namely: Deep Q-Learning (DQN) and double DQN (DDQN). These models are trained and tested at a 4-way intersection. Model adaptability and performance in different traffic scenarios are also measured and discussed in this paper.