Optimizing Autonomous Vehicle Performance Using Improved Proximal Policy Optimization.

Area of Science:

• Autonomous Driving
• Reinforcement Learning
• Robotics

Background:

• Autonomous vehicles require robust decision-making in dynamic urban environments.
• Proximal Policy Optimization (PPO) offers stability but suffers from limited exploration.
• Existing methods struggle with adaptability and exploring diverse strategies.

Purpose of the Study:

• To enhance the exploration capabilities of the PPO algorithm for autonomous driving.
• To improve the adaptability and reduce local minima entrapment in PPO.
• To develop a more stable and efficient learning mechanism for autonomous vehicles.

Main Methods:

• Integration of Lévy flight's chaotic exploration into the PPO algorithm, creating LFPPO.
• Real-time data collection from CARLA simulator (speed, location, traffic signals) processed via Apache Kafka.
• Comparative analysis of standard PPO and LFPPO performance in simulated urban traffic.

Main Results:

• LFPPO achieved a 99% success rate, significantly outperforming PPO's 81%.
• LFPPO reduced collision rates to 1%, compared to PPO's 19%.
• The enhanced algorithm demonstrated superior stability and higher rewards.

Conclusions:

• LFPPO effectively overcomes PPO's exploration limitations using Lévy flight and real-time data.
• The LFPPO algorithm offers enhanced safety and exploration for autonomous driving systems.
• This approach represents a significant advancement over current state-of-the-art methods.