Evaluating Deep Reinforcement Learning Methods to Develop an Intelligent Traffic Controller

In this study, deep reinforcement learning-based algorithms - Deep Q-Learning (DQN), Covariance Matrix Adaptation Evolution Strategy (CMA-ES), and Advantage Actor- Critic (A2C) - are tested to control vehicle traffic in Webots World Simulation. This work also uses a microscopic traffic simulator called Simulation of Urban Mobility (SUMO) to run traffic simulations. Ray, an open-source distributed computing platform, was used to accelerate the model's training, which uses a single dynamic execution engine to allow task-parallel and actor- based computations. The algorithms were evaluated and tested against fixed-time traffic control configurations and were proven more efficient. The results also show that A2C had the best performance among the learning-based approaches, while CMA- ES had the least in terms of total vehicle wait time.