Problematic Ordinal Input Categorical Output Neural Network

Hello everyone.! I'm fairly new to the ANN, but I am making quite an effort on that. I have a problem with ordinal input (numbers from 1-5 in each column, arrays are the observations (172) ) and categorical output (A-->10000 , B-->010000, C-->00100 etc) I am using patternet with 100 hidden neurons and

net.layers{1}.transferFcn = 'tansig';
net.layers{2}.transferFcn = 'logsig';

although I have given a shot to softmax and other transfer functions. Also I use

net.divideFcn = 'dividerand';
net.divideParam.trainRatio = 70/100;
net.divideParam.valRatio = 15/100;
net.divideParam.testRatio = 15/100;
net.trainFcn = 'trainscg';
net.trainParam.max_fail= 1000;

net.inputs{1}.processFcns = {'removeconstantrows','fixunknowns','mapminmax'};
net.outputs{2}.processFcns = {'removeconstantrows'};
net.trainParam.epochs=1000;
net.trainParam.min_grad=0;

this block in order to finalize my network. However, I have not achieved any greater accuracy the 52%.. Is there any specific suggestion.? Thank you everyone. I hope I was accurate enough.

Answer 1

[Greg Heath]

1. It looks like you can delete the above equations because they are defaults.

2. You definitely need to substantially reduce the No. of hidden nodes

 [ I N ] = size(input) % = ?
 [ O N ] = size(target)% = ?

 Ntrn = N -2*round(0.15*N)% = ?
 Ntrneq = Ntrn*O  % =? No. of training equations

For an I-H-O net, the number of unknown weights is

 Nw = (I+1)H+(H+1)*O

No. of unknowns is not larger than number of equations when

 H < Hub = (Ntrneq-O)/(I+O+1)

Using the code at

 help patternnet

and/or

 doc patternnet

where H = 10 (default)

find out how much of the average target variance the net can model

 MSE00 = mean((vartarget',1)); MSE reference
 NMSE  = mse(target-output)/MSE00 %Normalized MSE
 Rsq = 1-NMSE   ( > 0.99 is my goal)

Loop over i = 1:Ntrials = 10, Then check the 10 results to see if H can be decreased or should be increased.

For examples, search the NEWSGROUP and ANSWERS using

 patternnet greg

Hope this helps

Thank you for formally accepting my answer

Greg

Answer 2

[Andreas]

Thank you so much, I really appreciate your answer! I guess you mean

MSE00 = mean(var(target',1))

and I choose the setting with the highest Rsq. However, although my targets are formatted through the 1-out of-C philosophy, the outputs I get is not 0 or 1 vector, but there are also values inbetween (such as 0.012, 0.534 etc.) Is that correct? Should I post-process something?