Improving Transient Response of Model Reference Neuro-Controller via Constrained Optimization

A robust adaptation algorithm based on error normalization is introduced to update the weights of model reference neural network controller. Tracking error is normalized by a variable normalizing gain specified by solving a constrained optimization problem. The so-called piecewise quadratic cost function is proposed as the performance index to improve the transient response specifications. The conditions for robust convergence, saturation limit of actuators and maximum possible speed of response form the constraints of the problem in terms of the variable normalizing gain. Simulation results provided, demonstrate the improvements in transient behavior of control signal and output response obtained by the method, even in the presence of disturbances and parameter variations.