Optimizing Autonomous Driving with Advanced Reinforcement Learning: Evaluating DQN and PPO

Self-driving vehicles are considered one of the greatest technological achievements with the potential to restore the undeniable advantage of the automotive industry-safety, convenience, and economy. But creating 100% self-sufficient vehicles for actual scenery is a great challenge even to date. This research analyzes and finds out the efficiency of Deep Q-Network (DQN) and Proximal Policy Optimization (PPO) in enhancing autonomous driving (AD). In this study, the presented approach was implemented using the CARLA simulator, where various driving scenarios such as urban, highway, and obstacle avoidance driving scenarios were simulated. The efficiency of the algorithms was evaluated based on collision rate, time to destination, average speed, paths’ efficiency, and amount of time needed for computation. From the analysis of the results achieved, we can observe that indeed PPO surpasses DQN in all assessed aspects. A comparison of the collision rates also showed that PPO was overall safer because it had a lower collision rate The time to destination and average speeds also proved PPO to be a better driver because the numbers were smaller implying that PPO chauffeurs drove more efficiently. In addition, the results first demonstrated that PPO achieved higher path efficiency and indicated better path planning than the other algorithms. It can be noted that PPO involves a slightly higher computational cost; however, the performance benefits are worth the added cost. These statistics therefore establish PPO as a superior method to train real-world AV systems as compared to the conventional ones. Subsequent work will entail the development of these algorithms, their combination with other approaches of machine learning, and the problem of transferring learned policies into reality.