Deep reinforcement learning for traffic light optimization

Deep Reinforcement Learning has the potential of practically addressing one of the most pressing problems in road traffic management, namely that of traffic light optimization (TLO). The objective of the TLO problem is to set the timings (phase and duration) of traffic lights in order to minimize the overall travel time of the vehicles that traverse the road network. In this paper, we introduce a new reward function that is able to decrease travel time in a micro-simulator environment. More specifically, our reward function simultaneously takes the traffic flow and traffic delay into account in order to provide a solution to the TLO problem. We use both Deep Q-Learning and Policy Gradient approaches to solve the resulting reinforcement learning problem.