Deep Reinforcement Learning for Simulated Autonomous Vehicle Control

We investigate the use of Deep Q-Learning to control a simulated car via reinforcement learning. We start by implementing the approach of [5] ourselves, and then experimenting with various possible alterations to improve performance on our selected task. In particular, we experiment with various reward functions to induce specific driving behavior, double Q-learning, gradient update rules, and other hyperparameters. We find we are successfully able to train an agent to control the simulated car in JavaScript Racer [3] in some respects. Our agent successfully learned the turning operation, progressively gaining the ability to navigate larger sections of the simulated raceway without crashing. In obstacle avoidance, however, our agent faced challenges which we suspect are due to insufficient training time.