Keywords - Function groups - @ A B C D E F G H I J K L M N O P Q R S T U V W X Y Z

Library: nn
See also: committee nnrnet nnrpredict nnrinfo

Quantlet: neuronal
Description: This quantlet computes different networks of the form single layer feedforward perceptron. The quantlet can be used alone or in connection with the library ISTA. The standalone version also needs the parameter data. Just choose 0 for the input. It is possible to split the data in a training and a test set. The weight for the cases for the training of the net can be chosen, the numbers of hidden units with ``from, stepwidth, to'' and additional information concerning the weights of the units. Different optional parameters can be chosen to build the architektur of the network. The choice holds for every single net. The default values are chosen in order to solve a linear regression problem. The optional parameters constits of 8 values. Boolean values for linear output, entropy error function, log probability models and for skip connections (direkt links). The fifth values is the maximum value for the starting weights, the sixth is the weight decay, the seventh the maximum number of iterations and the the last value generates the output concerning the architekur of the net if it is equal to one. The output consits of the Error and MSE of the different nets (MSE for test and trainings data separately if chosen) and the R^2.

Reference(s):

Usage: proc(data)=neuronal(datain,x,y)
Input:
x n x p x d1 x ... x dn array
y n x 1 array
datain for use in ISTA: n x p x d1 x ... x dn array otherwise: 0
Output:
data for use in ISTA: n x p x d1 x ... x dn array otherwise: 0

Example:
library("nn")
datain=0
z=read("bostonh")
x=z[,1:13]
y=z[,14]
neuronal(datain,x,y)

Result:
Choose for the number of units in hidden layer 3, 1, 7, for the
weight decay 0.01 with 500 iterations and one for the output.
The architektur of the last network consists of:

A 13 - 7 - 1 network:
# weights     : 106
linear output : yes
error function: least squares
log prob model: no
skip links    : no
max. weight   : 0.70
decay         : 0.01
max. Iterat   : 500

Hidden Units  Error (train)  MSE (train)  MSE (test)  MSE (total)     R^2
_________________________________________________________________________
3      3645.1017      14.8117     25.6890      20.5083  0.7571
4      1392.2963       5.4750     15.6698      10.8141  0.8719
5      1390.0372       5.5051     20.0409      13.1177  0.8446
6      3880.2401      15.9997     23.4849      19.9198  0.7640
7      2285.1690       9.3947     16.6129      13.1750  0.8439



Author: K. Zanter, W. Haerdle, 19980331 license MD*Tech
(C) MD*TECH Method and Data Technologies, 05.02.2006