A neural-extended-sequential-filter-based model reference adaptive control system

A method for developing a model reference adaptive control system is proposed. It is assumed that the system being controlled is unknown except that it can be represented by a set of state equations with known dimension and that the measurement function vector is unknown except for its dimension. The unknown state function and measurement function are modeled by artificial neural networks and the unknown synaptic weights in the neural networks are determined by a neural extended sequential filter, as opposed to a neural extended Kalman filter, because the noises in the models are almost certainly not white. The adaptive controller is developed by driving a second neural extended sequential filter with appropriate signals from the model reference system and from the neural extended sequential filter which models the unknown system being controlled and its associated unknown measurement function. The technique presented in this paper is new and additional work is on-going to determine its practical value.