ADAPTIVE LEARNING NETWORKS OF MULTISTAGE FUZZY PRODUCTION RULES IN EXPERT-SYSTEM OF GRINDING CHARACTERISTICS

The cross validation technique offers a criterion to measure the degre e of approximation of a mathematical model. The advantage of the leavi ng-one-out error estimation and the k-fold cross validation is that al most all the available samples are used for training whereas all the s amples are used for testing. But because these techniques are computat ionally expensive, it has often been reserved for problems with small sample size. This paper discusses the validity of Akaike's AIC for sel ecting the number of layers in an adaptive learning network of GMDH ty pe whose partial descriptions are represented by Gaussian functions. I n numerical examples, several computer simulations of learning and com parisons of AIC with cross validation procedure are shown. The expert system for identification of grinding characteristics is discussed.