You are making this harder than it is. For the 1st time through all you have to specify is
1. Type of network
a fitnet for regression on curve-fitting
b. patternnet for classification or pattern-recognition
c. timedelaynet, narnet or narxnet for timeseries prediction
(fitnet and patternnet are special cases of feedforwardnet which never
has to be called explicitly)
2. input % size(input) = [ I N } for N I-dimensional Inputs
3. target % size(target) = [ O N ] for N O-dimensional Output targets
4. Ntrials: The number of candidate designs with different random data
divisions and random initial weights.
Then modify the code in the help and/or doc documentation. For example, use the commands
help fitnet
and
doc fitnet
to obtain code for fitnet. Put this code in a loop for ntrial = 1: Ntrials for different random data divisions and random initial weights.
This is the simplest code that I would use:
[ x t ] = simplefit_dataset;
vart = var(t,1)% Reference MSE
net = fitnet;
rng('default') % Needed for design duplication
for i = 1:Ntrials
net = configure(net,x,t);
[ net tr y e ] = train(net,x,t);
NMSE(i,1) = mse(e)/ vart; % Normalized MSE
end
output = NMSE
All 10 designs will be terrific because this is a simple data set. For other data I usually try to find the minimum number of hidden nodes (H, fitnet(H)) that will yield NMSE <= 0.01 ( i.e., the net successfully models 99% of the target variance)
Hope this helps.
Thank you for formally accepting my answer
Greg
