Approximate neural optimal control with reinforcement learning for a torsional pendulum device.

Abstract

A torsional pendulum device containing hyperbolic tangent input nonlinearities can be formulated as a nonaffine system. Unlike basic affine systems, the optimal feedback control of complex nonaffine plants is difficult but quite important. In this paper, the approximate optimal control design of continuous-time nonaffine nonlinear systems is investigated with the help of reinforcement learning. For addressing the learning algorithm conveniently, an effective pre-compensation technique is adopted to perform proper system transformation. Then, the integral policy iteration strategy is incorporated to relieve the demand of system dynamics. Moreover, the actor-critic structure is implemented by virtue of neural network approximators. Finally, the experimental verification for the proposed torsional pendulum plant is conducted after a learning process of 20 iterations and the stability performance with basic robustness guarantee can be observed during two case studies.