Neural Network .. the results are completely better when using (for loop) as compared when constructing and training each net individually.

Hi everyone; In my work I need to try different ANN topologies, preprocessing methods, activation functions, No. of layers, training functions, and so on. The purpose is to improve the final results of the regression model. I used for loop as in below; the results when using for loop is completely better than when constructing and training each network individually (I had tried to use one case of the output file for one network and do it manually, I found the result was very bad for the same network parameters as the RMSE was high). The question is whether my result is accurate or not? is it possible to use for a loop when you want to try hundreds of networks to find the best solution?

for i=1:4 % Network Topology loop FFANN,.. for typePrepro=1:7 % Preprocessing M loop zscore,...

    for layer=1:5                       % Max number of Network layers  
          NoNeurns=ones(1,layer);
          ActFn={};
          for fn = 1:layer
              ActFn{fn} = ActivationFn_New{1};
          end

          for layerLoop=1: layer            % Cycle for the chosen no. of layers
                for matrix=1:X    % Matrix for no of ActFun and NoLayers
                    ActFn{layerLoop}=M1{matrix};
                    if layerLoop ~= layer
                        NoNeurns(layerLoop)=M2(matrix);
                    else
                        NoNeurns(layerLoop)=1;
                    end

            for TrainFn=1:9                   %Training Functions (trainlm,...
            for TrainMthd=1:2                 % Training Methods (Adapt,...

Thanks

Answer 1

[Greg Heath]

I have designed hundreds of nets in the NEWSGROUP and ANSWERS using a double loop approach.

The outer loop is over number of hidden nodes,

 j = 0
 for h = Hmin:dH:Hmax
   j = j + 1
    ...
    net = fitnet(h);
    ...

the inner loop is over Ntrials subsets of random initial weights

    for i = 1:Ntrials
        net = configure(net,x,t);
        [ net tr y e ] = train( net, x ,t );
        ...
        NMSE(i,j) = mse(t-y)/mean(var(t',1));
    end
 end

Search words

Hmin Hmax Ntrials

should dredge up most of them.

Hope this helps.

Thank you for formally accepting my answer

Greg