Neural network based controllers for non-linear systems

In this paper, we present two approaches for the control of an inverted pendulum on a cart. First, we use a combination of a neural network and a simple algorithm, where the neural network is responsible for ranking the current states while the algorithm decides the control action (+10N or -10N) applied. Simulation results show that this method converges at a faster rate than previous researchers' schemes. In the next approach, the ability of artificial neural networks (ANN) to generate nonlinear mapping has been utilized to supplement classical control techniques, to perhaps achieve better performance. We present a hybrid controller consisting of a neural network and classical control technique. The neural network was trained to predict nonlinearities in the system. Having this prediction, a two term control law was designed where one term cancels the nonlinear effects: enabling us to use linear control theory (e.g. pole placement, optimal control, etc.) to obtain the second term. To check the applicability of our method, we tested this scheme on the bilinear model of a paper making machine as well. In these cases, our simulation studies revealed that improvements to the behavior of these systems could be achieved.<>