Abstract
Autonomous vehicles must make quick and accurate decisions to operate efficiently in complex and dynamic urban traffic environments, necessitating a reliable and stable learning mechanism. The proximal policy optimization (PPO) algorithm stands out among reinforcement learning (RL) methods for its consistent learning process, ensuring stable decisions under varying conditions while avoiding abrupt deviations during execution. However, the PPO algorithm often becomes trapped in a limited search space during policy updates, restricting its adaptability to environmental changes and alternative strategy exploration. To overcome this limitation, we integrated Lévy flight's chaotic and comprehensive exploration capabilities into the PPO algorithm. Our method helped the algorithm explore larger solution spaces and reduce the risk of getting stuck in local minima. In this study, we collected real-time data such as speed, acceleration, traffic sign positions, vehicle locations, traffic light statuses, and distances to surrounding objects from the CARLA simulator, processed via Apache Kafka. These data were analyzed by both the standard PPO and our novel Lévy flight-enhanced PPO (LFPPO) algorithm. While the PPO algorithm offers consistency, its limited exploration hampers adaptability. The LFPPO algorithm overcomes this by combining Lévy flight's chaotic exploration with Apache Kafka's real-time data streaming, an advancement absent in state-of-the-art methods. Tested in CARLA, the LFPPO algorithm achieved a 99% success rate compared to the PPO algorithm's 81%, demonstrating superior stability and rewards. These innovations enhance safety and RL exploration, with the LFPPO algorithm reducing collisions to 1% versus the PPO algorithm's 19%, advancing autonomous driving beyond existing techniques.