Actor-Critic Reinforcement Learning Control of Non-Strict Feedback Nonaffine Dynamic Systems

Abstract

The most focuses of the existing actor-critic reinforcement learning control (ARLC) are on dealing with continuous affine systems or discrete nonaffine systems. In this paper, I propose a new ARLC method for continuous nonaffine dynamic systems subject to unknown dynamics and external disturbances. A new input-to-state stable system is developed to establish an augmented dynamic system, from which I further get a strict-feedback affine model that is convenient for control designing based on a model transformation approach. The Nussbaum function is connected with a fuzzy approximation to devise an actor network whose tracking performance is further enhanced via strengthening signals generated by a fuzzy critic network. The stability of the closed-loop control system is guaranteed by the Lyapunov synthesis. Finally, the comparison simulation results are presented to verify the design.