The number of hidden neurons was therefore varied from two to

fore 13 sets of models were constructed and they employed two,

ur, till 14 hidden neurons. For each set of models, the AUC values

orded for the comparison. Figure 3.26(b) shows the result, where

seen that brnn preferred a model employing less hidden neurons

other two preferred a model employing more hidden neurons for

better performance, i.e., a greater AUC value.

dial basis function neural network

al basis function neural network algorithm (RBFNN) is an

n of the kernel density estimation approach to the classification

. The kernel density estimation approach is an unsupervised

learning approach, which does not involve class labels and is only

on one set of experimental data, which is normally the genotype

en analyse biological/medical data. When there are two

ntal data (one is the genotype data X and the other is the

e data y), the kernel density estimation approach can be extended

sification analysis problem to generate a model to map X to y.

nsion is called RBFNN and the basis function of RBFNN adopts

aped density distribution such as the Gaussian function.

ose there are N data points. A general format of a basis function

s defined as below, where x is a data point for estimating its

ܠ௡ is the n^th data points with known class information, ϑ is the

g parameter and ݂ሺܠሻ is the estimated density for x.

݂ሺܠሻൌ¹

ܰ^{෍݌ሺܠ|ܠ௡, ߴሻ}

ே

௡ୀଵ

(3.37)

earlier application of RBFNN is for nonlinear function

mation [Powell, 1977; Broomhead and Lowe, 1988]. Suppose

K basis functions. The output of a RBFNN is defined as below,