Neurofuzzy networks for online modelling and control with provable learning and stability conditions

This paper considers a wide class of basis associative memory networks and their learning and network conditioning for online modelling and control. It is shown that the networks parameter convergence rate, stability, and gradient noise all depend upon the condition number C(R) of the basis function autocorrelation function R. This analysis shows that for online modelling networks should be locally generalising and have condition number tending to unity.<>