Deep Learning vs. Discrete Reinforcement Learning for Adaptive Traffic Signal Control

The population in cities and demand for transportation continuously increases. Space, financial and environmental constraints do not allow for significant infrastructure expansion. Therefore, optimizing the efficiency of the infrastructure is becoming increasingly important. Wait time at traffic lights is a significant proportion of time spent travelling within cities. Time inefficiency of traffic lights is, therefore, a global concern. Adaptive traffic signal controllers aim to provide demand-responsive strategies to minimize motorists' delay and achieve higher throughputs at signalized intersections. With the advent of new sensory technologies and more intelligent control methods, we propose an adaptive traffic signal controller able to receive un-prepossessed high-dimensional sensory information and self-learn to minimize the intersection delay. We use (1) deep neural networks to operate directly on detailed sensory inputs and feed it into (2) a continuous reinforcement learning based optimal control agent. The integration of the two is known as deep reinforcement learning or deep learning for short. Using deep learning, we achieve two goals: (1) eliminate the need for handcrafting a feature extraction process such as determining queue lengths for instance, which is challenging and location specific, and (2) achieve better performance and faster training time compared to conventional discrete reinforcement learning approaches.