View article

This paper investigates learning control policies for traffic lights. We introduce a new reward function for the traffic light control problem, and propose the combination of the popular Deep Q-learning algorithm with a coordination algorithm for a scalable approach to controlling coordinating traffic lights, without requiring the simplifying assumptions made in earlier work. We show that this approach reduces travel times compared to earlier work on reinforcement learning methods for traffic light control and investigate possible causes of instability in the single-agent case.