Neurofuzzy velocity tracking control with reinforcement learning

Abstract

A control method of neurofuzzy controller with reinforcement learning is proposed to implement velocity tracking control of a vehicle manipulated by a robot driver. In neurofuzzy velocity tracking control, a neural network adjusts a fuzzy controller by fine-tuning the membership. The learning algorithm of the neural network is reinforcement learning, which is based on evaluating the system performance and giving credit for successful actions. After the designed controller is trained fully, the preliminary experiment is implemented to track the desired velocity. Experimental results show that the errors of velocity tracking meet the national standard. The maximum error usually appears at switch points of driving stages where displacement variety of acceleration pedal or brake pedal operated by the robot driver is larger. The error will reduce when displacement variety of the pedals becomes small except the near region of switch points. The robot driver embodies driving behavior of a skilled driver as operating the pedals, especially in the alternative switch process of the two pedals.