Traffic control with the use of multi- agent deep reinforcement learning

Abstract

In the modern society, traffic is a heated topic in everyday conversations and economics. As more and more traffic lights are given real-time access to traffic data from the intersections they are regulating, there is a potential for using multi-agent reinforcement learning to optimize traffic flow. In many cities, numerous intersections are often located close together and influencing each other in a complex pattern. By linking the intersections together for joint information processing based on deep artificial neural networks one can search for a solution that regulates the traffic flow in an optimized fashion. The approach chosen for our study is known as multi-agent deep reinforcement learning. For short we will just use the term multi-agent reinforcement learning. In the present thesis we generated two different models where the information shared between the intersections is different. The models were tested against a static model representing the classical way of controlling the intersections. From our results, we conclude that multi-agent reinforcement learning represents an interesting potential for improving the traffic flow in a network of intersections. This is especially the case when information is shared between the agents controlling the traffic lights at the various intersections. As this field of study advances, it is likely to expect that even better solutions requiring less computational power may be achieved.