drl-autonomous-driving-evaluation

Evaluation of Deep Reinforcement Learning methods in Autonomous Driving tasks

• Literature Review: Modern Deep Reinforcement Learning approaches for Autonomous Driving;
• Project Report: Evaluation of Deep Reinforcement Learning methods in Autonomous Driving.

Abstract

This project evaluates the effectiveness of three Deep Reinforcement Learning (DRL) methods, Deep Q-Networks (DQN), Proximal Policy Optimization (PPO), and Advantage Actor-Critic (A2C), in addressing autonomous driving challenges. Using a customizable simulation environment, the agents were trained and tested across four diverse driving scenarios: highway, roundabout, merge, and intersection. The analysis focused on both the training process (e.g., reward progression) and the post-training performance of the models, evaluating metrics such as total reward, collision rate, and driving behavior realism. Results showed that PPO generally achieved the best overall performance in terms of efficiency and realism. However, DQN delivered results that were often comparable or only slightly inferior to PPO, demonstrating robustness in various scenarios. A2C, while effective in some cases, struggled with consistency and adaptability.