Asynchronous n-step Q-learning adaptive traffic signal control

Ensuring transportation systems are efficient is a priority for modern society. Intersection traffic signal control can be modeled as a sequential decision-making problem. To learn how to make the best decisions, we apply reinforcement learning techniques with function approximation to train an adaptive traffic signal controller. We use the asynchronous n-step Q-learning algorithm with a two hidden layer artificial neural network as our reinforcement learning agent. A dynamic, stochastic rush hour simulation is developed to test the agent's performance. Compared against traditional loop detector actuated and linear Q-learning traffic signal control methods, our reinforcement learning model develops a superior control policy, reducing mean total delay by up 40% without compromising throughput. However, we find our proposed model slightly increases delay for left turning vehicles compared to the actuated controller, as a consequence of the reward function, highlighting the need for an appropriate reward function which truly develops the desired policy.