Hello, I am new to ANN. I wanted to develop a network (fitnet) and used the following code in MATLAB (6 inputs and 1 output).

But, the overall R=0.56. Are there any bugs in the code? How can I further improve this network? Is it possible to use ensemble techniques or advanced initialization techniques to this type of network (ex: glorot)?

Import the data

x=myModel(:,1:6);
y=myModel(:,7);
l=length(x);

Normalize the inputs

for k=1:6
    xtnn(:,k)=(x(:,k)-min(x(:,k)))/(max(x(:,k))-min(x(:,k)));
end
for f=1
    ytnn(:,f)=(y(:,f)-min(y(:,f)))/(max(y(:,f))-min(y(:,f)));
end

Train the ANN

trainFcn='trainlm' 
xtn=xtnn'; % transpose of the input matrix
ytn=ytnn'; %transpose of the output matrix

hiddenLayerSize=(23);

netM=fitnet(hiddenLayerSize,trainFcn);

netM.divideFcn='divideind';

K-Fold Cross Validation

h=10;
c = crossvalind('KFold',l,10);
rng('default');

for i=1:h
    
    netM=configure(netM,xtn,ytn);
    
    vlInd=(c == i);
    trInd= ~vlInd;
    trInx=find(trInd);
    vlInx=find(vlInd);
    
    netM.divideParam.trainInd=trInx;
    netM.divideParam.valInd=vlInx;
    netM.trainParam.lr=0.01; 
    netM.trainParam.epochs=10000; 
    netM.performParam.regularization = 0.01;
    netM.performFcn='mse'; 
    netM.layers{1}.transferFcn='tansig';
    netM.layers{2}.transferFcn='purelin';
    netM.layers{1}.initFcn='initnw';
    netM.layers{2}.initFcn='initnw';
    
    tr.trainInd = trInx;
    tr.valInd = vlInx;
    
    net.initFcn='initlay';
    netM=init(netM);
    [netM,tr]=train(netM,xtn,ytn); 
end

Output = sim(netM,xtn,[],[],ytn)
plotperf(tr)
trOutput=Output(trInx)
valOutput=Output(vlInx)
trTarget=ytn(trInx)
valTarget=ytn(vlInx)
plotregression(trTarget, trOutput, 'Train', valTarget, valOutput, 'Validation', ytn, Output, 'All')
tr.best_epoch

Obtaining the weights and bias matrices

w1=netM.IW{1,1};
w1'
w2=netM.LW{2,1};
w2'
b1=netM.b{1}
b2=netM.b{2}

Unnormalized Output

Output_I=Output';
Output_R=((Output_I(:,f))*(max(y(:,f))-min(y(:,f))))+(min(y(:,f)))

Performance of the network

perf=mse(netM,ytn,AN4)
save netM