System linearization with guaranteed stability using norm-bounded neural networks

A new method for linearizing nonlinear plants with neural networks resulting in robustly stable closed-loop systems is presented. The class of plants considered constitutes a set of unknown but invertible nonlinear systems. In this method, neural network outputs are treated as parametric uncertainty and are combined with other plant uncertainties so that a robust controller can be designed. An algorithm for confining the network´s output to be less than a given bound is presented. We demonstrate the effectiveness of using a linear inner loop feedback to reduce the size of a neural network for robust control purposes. The method was successfully applied to the inverted pendulum problem and simulation results indicate that our approach performed very well