A stability analysis of direct neuro-fuzzy controllers

Within the framework of input/output stability we develop an algorithm for testing the stability of a given direct neuro-fuzzy controller in a feedback situation, dependant only on the functional gain of the plant to be controlled. A series of geometrical descriptions are then developed for the set of weights for which the feedback loop remains stable. These are then exploited to give adaptive mechanisms, where training can be achieved on line, whilst maintaining stability of the feedback loop.