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: neuronal2 nnrnet2 nnrpredict2 nnrinfo

Quantlet: committee2
Description: This quantlet computes a committee of networks with nets 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. The number of nets to build the committee can be chosen. The data will be splitted with this number to build the different datasets. The weight for the cases for the training of the net can be chosen, the numbers of hidden units 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 single nets and for all cases. Additionally the R^2 for the average of the nets and the R^2 of the committee are shown.

Reference(s):

Usage: proc(data)=committee2(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]
committee2(datain,x,y)

Result:
Choose 4 single nets to build teh committee.
For the number of units in hidden layer choose 3, for the
weight decay 0.01 with 500 iterations and one for the output.
The architektur of the nets in the committee consists of:

A 13 - 3 - 1 network:
# weights     : 46
linear output : yes
error function: least squares
log prob model: no
skip links    : no
max. weight   : 0.70
decay         : 0.01
max. Iterat   : 500

committee networks
net  Error (single net) MSE (single net)  MSE (all cases)
_________________________________________________________
1           4126.8667          10.4246           0.4772
2           4136.4091          11.0771          29.9052
3           3854.8416          10.4423          18.9735
4           3634.2495           9.2975           8.1649
_________________________________________________________
MSE of networks  14.3802      R^2: 0.8297
MSE due to committee  2.4619  R^2: 0.9708



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