Nowadays, driver-assistance systems have been extensively researched and implemented to increase driving safety. There are many applications derived from this concept, such as Adaptive Cruise Control (ACC) system, which is now used in some automobile for safety driving. In the ACC system, with the help of radars or other sensors equipped for detecting the distance and relative speed to the target vehicle, the control system will help the driver keep a safe distance and relative speed. ACC can not only liberate the driver from frequent speeding up and slowing down but also reduce the drivers’ mental stress. However, traditional control strategy can hardly fulfill the human-like control requirement, even function unsatisfied under complex environments. As a result, to development a simple, easy and effective controller is an urgent need. On the other hand, reinforcement learning (RL) has aroused abroad attention in the research areas because of the self-learning property. Supervised reinforcement learning (SRL) has both the merits of RL and supervised learning (SL). It is very suitable for the ACC problem with human-like control requirement. Adaptive dynamic programming (ADP) can be deemed as a higher level RL. It uses a neural network to approximate the state value that will be gained through iterating in dynamic programming to solve the curse of dimensionality problem. In this paper, we analyze the driving mode, construct the upper control model for ACC and introduce RL, SRL as well as ADP. At last, we propose Supervised adaptive dynamic programming (SADP) based on ADP and SL. We design the SRL, SADP and hybrid PID controllers to realize the human-like control, and then apply these controllers to different simulation scenarios. The results show that the SADP controller outperforms the SRL controller. The SADP controller also outperforms the hybrid PID controller slightly because the speed and distance control in the SADP controller are smoother than that in the hybrid PID controller. The acceleration in the SADP controller is smaller and smoother than that in the hybrid PID controller, and the time to collision (TTC) in the SADP controller is as good as that in the hybrid PID controller. It can be concluded that the SADP controller not only is robust, but also has enough accuracy in the control performance. The SADP controller provides us with a feasible and effective way in the human-like ACC system.
修改评论