A general structure for neurocontrol of dynamical systems

A novel neurocontrol configuration consisting of three multilayer neural networks (MNNs) is proposed for nonlinear MIMO plants. Two MNNs, one with feedback inputs and the other without feedback, are used to perform plant identification based on a generalized state variable model for plant representation. The remaining MNN acts as a feedforward controller, by taking advantage of the state information generated as a by-product of the identification process. No hypothesis about the plant invertibility is assumed, overcoming the main drawback of most neurocontrol schemes suggested in the literature so far. The proposed configuration is very general, and only a few parameters must be user-specified. Training algorithms are provided in pseudocode and preliminary results attest the practicability and effectiveness of the proposed scheme.<>