The most important objective is to obtain an acceptable ratio of mean-square error to mean-target-variance for RANDOM subsets of NONTRAINING data. I find the best way to do this is to try to find the smallest value of H (No. of hidden layer nodes) that will yield
mse(ytst-ttst)/var(ttst,1) <= 0.01
i. e., 99% of a random test target subset variance is successfully modeled.
In order to do this I use the default 70/15/15 trn/val/tst data division and the training data goals
MSEgoal = max( 0, 0.01*(Ntrneq-Nw)*var(ttrn,0)/Ntrneq )
MinGrad = MSEgoal/100
net.trainParam.goal = MSEgoal ;
net.trainParam.min_grad = MinGrad ;
For an I-H-O node topology obtained from N pairs of I-dimensional inputs and O-dimensional targets
[ I N ] = size(input) % [ 85 130 ]
[ O N ] = size(target)% [ 26 130 ]
Ntrn = N-2*round(0.15*N) % 90 training examples
Ntrneq = Ntrn*O % 2340 training equations
% Nw = (I+1)*H+(H+1)*O % Number of unknown weights
% For a robust design desire Ntrneq >> Nw or
% H << Hub = -1+ceil( (Ntrneq-O) / (I+O+1)) % 20
Hmax = 20
dH = 2
Hmin = 1
Ntrials = 10
Design many nets and choose the best using the validation set performance.
for h = Hmin:dH:Hmax
....
for i = 1: Ntrials
...
end
end
I have posted many, many examples on the NEWSGROUP and ANSWERS, Search on
greg Ntrials
Hope this helps.
Thank you for formally accepting my answer
Greg
