P. Yee et S. Haykin, A dynamic regularized radial basis function network for nonlinear, nonstationary time series prediction, IEEE SIGNAL, 47(9), 1999, pp. 2503-2521
In this paper, constructive approximation theorems are given which show tha
t under certain conditions, the standard Nadaraya-Watson regression estimat
e (NWRE) can be considered a specially regularized form of radial basis fun
ction networks (RBFN's), From this and another related result, we deduce th
at regularized RBFN's are m.s. consistent, like the NWRF for the one-step-a
head prediction of Markovian nonstationary, nonlinear autoregressive time s
eries generated by i.i.d. noise processes, Additionally, choosing the regul
arization parameter to be asymptotically optimal gives regularized RBFN's t
he advantage of asymptotically realizing minimum m.s. prediction error. Two
update algorithms (one with augmented networks/infinite memory and the oth
er with fixed-size networks/finite memory) are then proposed to deal with n
onstationarity induced by time-varying regression functions. For the latter
algorithm, tests on several phonetically balanced male and female speech s
amples show an average 2,2-dB improvement in the predicted signal/noise (er
ror) ratio over corresponding adaptive linear predictors using the exponent
ially-weighted RLS algorithm. Further RLS filtering of the predictions from
an ensemble of three such RBFN's combined with the usual autoregressive in
puts increases the improvement to 4.2 dB, on average, over the linear predi
ctors.