Traffic Light Control with Reinforcement Learning

Devanshi Malaviya, CIS494: Research in Computer Information SystemsFaculty Mentor(s): Professor Sarbani Banerjee, Computer Information Systems The rapid increase in automobiles in the past few years has led to traffic congestions all over the world. This forces drivers to sit idly in their cars wasting time and fuel. Current traffic light control policies are not optimized, which leads to people waiting in their cars for nonexistent traffic and more extended travel time than necessary. In the United States, motorists spend an average of nearly 100 hours in traffic congestions per year. The current research project focuses on reinforcement learning to optimize traffic flow to reduce travel time of drivers. It can be done by building an environment where every intersection has knowledge about the number of the vehicles and their speed as they approach the intersection. Simulation of Urban Mobility (SUMO) is used to build a traffic simulator. Reinforcement learning works on state and action policies which allow traffic lights to make optimized decisions based on their current state. It will balance the exploration and exploitation to make sure that the model is not overfitting and every lane is given importance according to how busy it is. For every state, it receives a reward if it reduces travel time, and the goal of the model is to collect as many rewards as possible. Therefore, the project will conclude by attempting to obtain the most optimized simulation. The Python packages used in the project are Keras, Tensorflow and OpenAI.https://digitalcommons.buffalostate.edu/srcc-sp21-compeng/1009/thumbnail.jp