Intelligent Traffic Control System using Deep Reinforcement Learning

Abstract

In this paper, we propose a deep reinforcement learning based traffic signal controller. We use the recently developed Distributional Reinforcement Learning with Quantile Regression (QR-DQN) algorithm to design a risk-sensitive approach to traffic signal control. A neural network is used to estimate the value distribution of state-action pairs. A novel control policy that gives variable weightage to the risk of an action depending on the congestion state of the system, effectively minimizes congestion in the network. Our results show that our algorithm outperforms conventional approaches and also classic RL based ones.