Applying Hierarchal Clusters on Deep Reinforcement Learning Controlled Traffic Network

Abstract

Traffic congestions is a crucial problem affecting cities around the globe and they are only getting worse as the number of vehicles tends to increase significantly. Traffic signal controllers are considered as the most important mechanism to control traffic, specifically at intersections, the field of Machine Learning introduces advanced techniques which can be applied to provide more flexibility and adaptiveness to traffic control techniques. Efficient traffic controllers can be designed using a reinforcement learning (RL) approach but major problems of following RL approach are, exponential growth in the state and action spaces and the need for coordination. We use real traffic data of 65 intersection of the city of Ottawa to build our simulations and show that, clustering the network using hierarchal techniques has a great potential in reducing the state-action pair significantly and enhance overall traffic performance.