backpropogation ,Multilayer perceptron,neural network

dear sir,

i m attaching a matlab code in which i tried to train the network using Feed forward Backpropogation.Here i m finding difficulty in in instruction

net_FFBP = createNet(inputsize, mimax, hneurons, fcnCELL, initflag, trainalgo, paramatrix, sameWEIGHT);

can i get how to save parameters in net_FFBP.I have attached the code below

function TrainingNet

load Feature.txt;    %load the features

FeatureS = Feature';    %Convert to column array

load Outtype.txt;     %load output type

OuttypeS = Outtype';

inputsize = size(FeatureS, 1);
min_data = min(min(FeatureS));
max_data = max(max(FeatureS));
mimax = [min_data max_data];
hneurons = 2000;
%initialize parameters for creating the MLP.
fcnCELL = {'logsig' 'logsig'};
initflag = [0 1];
trainalgo = 'gdm';
paramatrix = [10000 50 0.9 0.6]; % epochs = 100, show = 50, learning rate = 0.9, momentum term = 0.6 
sameWEIGHT = [];

net_FFBP = creteNet(inputsize, mimax, hneurons, fcnCELL, initflag, trainalgo, paramatrix, sameWEIGHT);

net_FFBP = newff(FeatureS, OuttypeS, 39);

[net_FFBP] = train(net_FFBP, FeatureS, OuttypeS);

save net_FFBP net_FFBP;

disp('Done: Training Network');

Answer 1

[Greg Heath]

% function TrainingNet

% load Feature.txt; %load the features

% FeatureS = Feature'; %Convert to column array

% load Outtype.txt; %load output type

% OuttypeS = Outtype';

[I N ] = size(FeatureS)

[O N ] = size(Outtypes)

minmaxF = minmax(FeatureS) % Is a matrix [I 2]

Neq = N*O % Number of training equations

% I-H-O node topology

% Nw = (I+1)*H+(H+1)*O % Number of unknown weights

% Want Neq >> Nw or % H << Hub

Hub = (Neq-O)/(I+O+1) % Neq = Nw

r = 10 % Neq > r*Nw, ~2 < r < ~30

H = floor((Neq/r-O)/(I+O+1))

How did you get H = 2000 ???

% %initialize parameters for creating the MLP.

% fcnCELL = {'logsig' 'logsig'};

% initflag = [0 1];

What does initflag do?

% trainalgo = 'gdm';

% paramatrix = [10000 50 0.9 0.6]; % epochs = 100, show = 50,

100 or 10,000?

                                   % learning rate = 0.9, momentum term = 0.6

% sameWEIGHT = [];

I suggest first using the defaults in NEWFF

% net_FFBP = creteNet(inputsize, mimax, hneurons, fcnCELL, initflag, trainalgo, paramatrix, sameWEIGHT);

Is this supposed to be a replacement for NEWFF and net.Param.* ??

% net_FFBP = newff(FeatureS, OuttypeS, 39);

Now H = 39 ??

% [net_FFBP] = train(net_FFBP, FeatureS, OuttypeS);

% save net_FFBP net_FFBP;

% disp('Done: Training Network');

What is your question ??

Greg