A Personalized Federated Imitation Learning Algorithm for Autonomous Driving

Currently, end-to-end autonomous driving systems that employ imitation learning effectively learn and optimize driving strategies by integrating the driving behaviors of human experts with modern deep learning techniques. However, challenges such as scene diversity, data security, and training time must still be addressed to develop a model that is applicable across various traffic scenarios while maintaining high accuracy. To tackle these issues, this paper proposes a personalized feder-ated imitation learning algorithm that aggregates locally trained imitation learning decision models from vehicles operating in different scenarios through a distributed training approach, thereby enhancing training efficiency and model accuracy. Specifically, considering the variability in communication link quality and potential disconnections caused by the mobility of vehicles in a connected vehicle network, we introduce a personalized user selection algorithm. Building on this foundation, we employ a federated imitation learning method to efficiently and rapidly train a driving decision model with comparable performance while safeguarding data privacy. Extensive simulation results confirm the superiority of the proposed algorithm in terms of training speed and model accuracy.