Improving the Performance of Autonomous Driving through Deep Reinforcement Learning

Abstract

Reinforcement learning (RL) is revolutionizing the artificial intelligence (AI) domain and significantly aiding in building autonomous systems with a higher level comprehension of the world as we observe it. Deep learning (DL) facilitates RL to scale and resolve previously intractable problems, for instance, allowing supervision principles designed for robots to be acquired directly from visual data, developing video game proficiency from pixel-level information, etc. Recent research shows that RL algorithms help represent problems dealing with high-dimensional, unprocessed data input and can have successful applications in computer vision, pattern identification, natural language analysis, and speech parsing. This research paper focuses on training a simulation model of a car to navigate autonomously on a racetrack using RL. The study explores several fundamental algorithms in Deep RL, namely Proximal Policy Optimization (PPO), Deep Q-network (DQN), and Deep Deterministic Policy Gradient (DDPG). The paper documents a comparative analysis of these three prominent algorithms—based on their speed, accuracy, and overall performance. After a thorough evaluation, the research indicates that the DQN surpassed the other existing algorithms. This study further examined the performance of the DQN with and without ε-decay and observed that the DQN with ε-decay is better suited for our objective and is significantly more stable than its non ε-decay counterpart. The findings from this research could assist in improving the performance and stability of autonomous vehicles using the DQN with ε -decay. It concludes by discussing the fine-tuning of the model for future real-world applications and the potential research areas within the field of autonomous driving.