Decision Making for Autonomous Driving in a Virtual Highway Environment based on Generative Adversarial Imitation Learning

Autonomous driving has gained significant attention as a rapidly advancing research field in recent years. This paper addresses the challenge of efficient highway driving by formulating it as a Markov Decision Process (MDP) and leveraging Reinforcement Learning (RL) techniques. To tackle the decision-making problem, Generative Adversarial Imitation Learning (GAIL) is employed to imitate expert behavior. In this study, we acquired expert data by running an agent within the simulation, enabling us to gather a larger volume of data. Then, the expert agent is trained by Deep Q-Network (DQN) using reward shaping. Our contribution in this study involves the application of Curriculum Learning (CL) specifically to highway scenarios, gradually increasing the complexity of the traffic environment to enhance the training process. In order to achieve a behavior comparable to that of a human driver, we proposed the use of the GAIL approach specifically for highway scenarios, aiming to increase the diversity of state and action pairs. Our experiments successfully demonstrated the effectiveness of this approach, as the autonomous driving agent effectively imitated expert behavior and achieved outstanding collision-free performance on highways.