Robot ℋ 2 Neural Control Using Reinforcement Learning

Neural network is one of the most widely used approximators for reinforcement learning (RL) methods and model‐learning approaches. This chapter discusses H₂ neural control based on recurrent neural networks and RL for robots, which has better performance than the standard linear quadratic regulator control and also accelerates the learning process. The controllers are designed in both discrete time and continuous time. Control of discrete‐time systems is becoming important in recent years since almost all of the control schemes are implemented on digital devices. The neural RL has good mean squared error results by taking into account measures of the position and velocity states and the modeling error. The chapter presents stability and convergence using Lyapunov stability theory and the contraction property. Simulations results show that the RL approach presents optimal and robust performances without knowledge of the system dynamics and large learning time.