Adaptive neural network control with frequency-shaped optimal output feedback

The concept of a neural-network-based adaptive optimal control scheme for dynamic systems is investigated. The neural network (NN) is used to learn and generate the coefficients of several predesigned discrete-time controllers as a function of external parameters. The result yields a controller whose frequency characteristics are adaptive to internal and external factors. The proposed method relies on the ability of an NN to learn and smoothly interpolate the controller coefficients. Linear optimal output feedback controllers are designed using a class of frequency-shaped quadratic performance measures and are then used to train the NN. These frequency-shaped controllers provide optimum control performance for different bands of input frequencies. The resulting NN-based adaptive optimal controllers can be utilized for extremely complex systems in which conventional nonlinear/adaptive control techniques are forbidding or of no use. The authors discuss an application of the theory and methodology to an adaptive optimal tracking system that is subjected to a wide range of external inputs