Human-like Adaptive Cruise Control System Based on Fuzzy Reinforcement Learning

With the great development of artificial intelligence, it is possible to build an autonomous driving system that has truly driving intelligence. As one of the basic functions of autonomous driving, adaptive cruise control (ACC) needs to further improve safety and ride comfort on the existing basis. Recently, reinforcement learning (RL) shows great potential on building ACC systems. However, low learning efficiency and slow convergence speed limit its practical applications. This paper proposes an ACC system combined with semantic information based on fuzzy RL, which mainly includes the following two contributions: 1) a fuzzy RL algorithm with a new update method that is suitable for vehicle following; 2) a multivariate reward function is proposed for the ACC, taking into account both the target reward and the process reward, to ensure safety and driving stability of the vehicle while improving the training efficiency. The experimental results show that our algorithm greatly improves the learning speed of Q- Learning under the vehicle following tasks. Unlike the traditional ACC strategy, the learned control strategy is highly similar to that of human drivers. While accurately tracking the target speed, the ride comfort and safety of driving process are fully considered, which reflects a high degree of driving intelligence.