MIMO Neural Network Performance

Hi, I am developing MIMO Neural Network for Integrated Model-Predictive Control.

I have some problems in developing Model.

First, I can't optimize all of outputs from MIMO Neural Network.

Though I normalized all targets and inputs, MIMO Neural Network's Each Performance does not get better.

even, I made initialize function of all layerWeights and InputWeights, Biases same.

and I choose network performance parameter as 'standard'.

Second, so I think the way to copy weights and biases from MISO Neural net to MIMO Neural net.

but it has difference performance of each target. (after copying, I didn't train MIMO Neural net.)

is there anyone who can solve this problems?

this is sample code1 .

rng(0,'twister')
x1 = [linspace(1,10,100);linspace(11,20,100)];
x2 = [linspace(11,20,100);linspace(21,30,100)];
t1 = linspace(101,200,100);
t2 = linspace(201,300,100)*100;

for i = 1:2
   x1(i,:) = (x1(i,:)-min(x1(i,:)))/(max(x1(i,:))-min(x1(i,:)));
   x2(i,:) = (x2(i,:)-min(x2(i,:)))/(max(x2(i,:))-min(x2(i,:)));
end

q = linspace(0.001,0.01,10);
r = linspace(0.001,0.1,10);

for i = 1:10
    for s =1:10
        net=fitnet(10,'trainscg');
        net.performParam.normalization ='standard';
        net.numLayers=4;
        net.layerConnect(4,3)=1;
        net.layers{3}.transferFcn = 'tansig';
        net.layers{3}.initFcn = 'initnw';
        net.layers{4}.initFcn = 'initnw';

        net.layers{3}.size = 10;
        net.biasConnect(3)=1;
        net.biasConnect(4)=1;
        net.numInputs=2;
        net.outputConnect(4)=1;
        net.inputConnect(3,2)=1;
        net.inputConnect(3,1)=1;


        net.divideFcn ='';
        net.inputs{1}.processFcns={};
        net.inputs{2}.processFcns={};
        net.outputs{2}.processFcns={};
        net.outputs{4}.processFcns={};

        net.layerWeights{4,3}.learnFcn ='learngdm';
        net.inputWeights{3,2}.learnFcn ='learngdm';
        net.inputWeights{3,1}.learnFcn ='learngdm';

        net.inputWeights{1,1}.learnParam.lr =q(i);
        net.inputWeights{1,1}.learnParam.mc =r(s);
        net.inputWeights{3,2}.learnParam.lr =q(i);
        net.inputWeights{3,2}.learnParam.mc =r(s);

        net.layerWeights{4,3}.learnParam.lr =q(i);
        net.layerWeights{4,3}.learnParam.mc =r(s);
        net.layerWeights{2,1}.learnParam.lr =q(i);
        net.layerWeights{2,1}.learnParam.mc =r(s);

        net.biases{3}.learnFcn = 'learngdm';
        net.biases{4}.learnFcn = 'learngdm';

        net.biases{1}.learnParam.lr =q(i);
        net.biases{1}.learnParam.mc =r(s);
        net.biases{2}.learnParam.lr =q(i);
        net.biases{2}.learnParam.mc =r(s);

        net.biases{3}.learnParam.lr =q(i);
        net.biases{4}.learnParam.lr =q(i);
        net.biases{3}.learnParam.mc =r(s);
        net.biases{4}.learnParam.mc =r(s);


        net=train(net,{x1;x2},{t1;t2});

        p = net({x1;x2});

        first_error1(i,s) = sqrt(mean((t1-p{1,:}).^2))/mean(t1)*100;
        first_error2(i,s) = sqrt(mean((t2-p{2,:}).^2))/mean(t2)*100;
        disp(first_error1(i,s))
        disp(first_error2(i,s))
    end
end

the sample code2 is

rng(0,'twister')
x1 = [linspace(1,10,100);linspace(11,20,100)];
x2 = [linspace(11,20,100);linspace(21,30,100)];
t1 = linspace(101,200,100);
t2 = linspace(201,300,100)*100;

for i = 1:2
   x1(i,:) = (x1(i,:)-min(x1(i,:)))/(max(x1(i,:))-min(x1(i,:)));
   x2(i,:) = (x2(i,:)-min(x2(i,:)))/(max(x2(i,:))-min(x2(i,:)));
end

net1 = fitnet(10,'trainscg');
net1.divideFcn ='';
net1.inputs{1}.processFcns={};
net1.outputs{2}.processFcns={};

net2 = fitnet(10,'trainscg');
net2.numInputs=2;
net2.inputConnect(1,1) = 1;
net2.inputConnect(1,2) = 1;
net2.divideFcn ='';
net2.inputs{1}.processFcns={};
net2.inputs{2}.processFcns={};
net2.outputs{2}.processFcns={};

net1 = train(net1,x1,t1);
net2 = train(net2,{x1;x2},t2);

p = net1(x1);
net1_error1 = sqrt(mean((t1-p).^2))/mean(t1)*100;

p = net2({x1;x2});
net2_error2 = sqrt(mean((t2-cell2mat(p)).^2))/mean(t2)*100;


net=fitnet(10,'trainscg');
net.trainParam.showWindow=false;
net.performParam.normalization ='standard';
net.numLayers=4;
net.layerConnect(4,3)=1;
net.layers{3}.transferFcn = 'tansig';
net.layers{3}.initFcn = 'initnw';
net.layers{4}.initFcn = 'initnw';
net.layers{3}.size = 10;
net.biasConnect(3)=1;
net.biasConnect(4)=1;
net.numInputs=2;
net.outputConnect(4)=1;
net.inputConnect(3,1)=1;
net.inputConnect(3,2)=1;

net.divideFcn ='';
net.inputs{1}.processFcns={};
net.inputs{2}.processFcns={};
net.outputs{2}.processFcns={};
net.outputs{4}.processFcns={};

net.layerWeights{4,3}.learnFcn ='learngdm';
net.inputWeights{3,2}.learnFcn ='learngdm';

net.inputWeights{3,2}.learnParam.lr =0.01;
net.inputWeights{3,2}.learnParam.mc =0.09;
net.layerWeights{4,3}.learnParam.lr =0.01;
net.layerWeights{4,3}.learnParam.mc =0.09;

net.biases{3}.learnFcn = 'learngdm';
net.biases{4}.learnFcn = 'learngdm';

net.biases{1}.learnParam.lr =0.01;
net.biases{1}.learnParam.mc =0.09;
net.biases{2}.learnParam.lr =0.01;
net.biases{2}.learnParam.mc =0.09;

net.biases{3}.learnParam.lr =0.01;
net.biases{4}.learnParam.lr =0.01;
net.biases{3}.learnParam.mc =0.09;
net.biases{4}.learnParam.mc =0.09;

net = configure(net,{x1;x2},{t2;t1}); % During Output Connecting, Output's order gets different

net.IW{1,1}=net1.IW{1,1};
net.IW{3,1}=net2.IW{1,1};
net.IW{3,2}=net2.IW{1,2};
net.LW{2,1} = net1.LW{2,1};
net.LW{4,3} = net2.LW{2,1};

net.b(1) = net1.b(1);
net.b(2) = net1.b(2);
net.b(3) = net2.b(1);
net.b(4) = net2.b(2);

p = net({x1;x2});
net_error1 = sqrt(mean((t1-p{2,:}).^2))/mean(t2)*100;
net_error2 = sqrt(mean((t2-p{1,:}).^2))/mean(t2)*100;