Advanced Deep Reinforcement Learning Strategies for Enhanced Autonomous Vehicle Navigation Systems

Abstract

The objective of this research is to investigate more sophisticated techniques in Deep Reinforcement Learning (DRL) for the purpose of improving navigation systems for autonomous vehicles. The results of our research indicate that the success rates of DRL algorithms are encouraging: DQN (75.2%), PPO (82.6%), and TRPO (78.1%). The success of navigation is highly influenced by optimisation factors such as parameter tuning (87.4%), exploration-exploitation balance (85.1%), and reward shaping (88.9%). The generalisation score for transfer learning is rather good (92.3%), while the safety modules are designed to make the system safer. These findings highlight the significance of algorithm selection, parameter optimisation, and advanced techniques in the process of optimising navigation for autonomous vehicles, hence establishing essential benchmarks for future developments.