my neural network is giving same output for all inputs...do you have any idea why?

net=network(8,3,[1;1;1],[1 1 1 1 1 1 1 1;0 0 0 0 0 0 0 0;0 0 0 0 0 0 0 0],[0 0 0;1 0 0;0 1 0],[0 0 1]); net.layers{1}.transferFcn='logsig'; net.layers{2}.transferFcn='logsig'; net.layers{3}.transferFcn='logsig'; net.layers{2}.dimensions=10; net.trainFcn='traingd'; net.trainparam.min_grad=0.00001; net.trainparam.epochs=10000; net.trainparam.lr=0.3; net.trainparam.goal=0.0001; net=init(net); net.layers{1}.initFcn='initwb'; net.layers{2}.initFcn='initwb'; net.biases{1,1}.initFcn='rands'; net.biases{2,1}.initFcn='rands'; i=load('input.txt'); t=load('target.txt'); i=i'; t=t'; in=zeros(8,53); %normalized input tn=zeros(1,53); %normalized target

for r=1:8 %normalization of input min=i(r,1); max=i(r,1); for c=2:53 if i(r,c)<min min=i(r,c); end if i(r,c)>max max=i(r,c); end end for c=1:53 in(r,c)=0.1+(0.8*(i(r,c)-min)/(max-min)); end end

min=t(1); %normalization of target max=t(1); for c=2:53 if t(1,c)<min min=t(1,c); end if t(1,c)>max max=t(1,c); end end for c=1:53 tn(1,c)=0.1+(0.8*(t(1,c)-min)/(max-min)); end

net.divideFcn='divideblock'; net.divideParam.trainRatio = 0.85; net.divideParam.valRatio = 0.05; net.divideParam.testRatio = 0.1; net.performFcn='mse'; [net,tr]=train(net,in,tn); y=sim(net,in);

[Jan]

Please read and apply http://www.mathworks.com/matlabcentral/answers/13205-tutorial-how-to-format-your-question-with-markup.

[Greg Heath]

1. Why would you post a long code that will not run when cut and pasted into the command line because there is no sample data???

2. NEVER use MATLAB function names for your own variables (e.g., max and min)

2. When beginning to write a program it is smart to try to use all of the defaults of the functions and use MATLAB data that is most similar to yours.

   help nndata

3. Once that runs you can begin to modify it to fit your original problem.

4. Cut and paste the program to make sure it runs or to obtain the error messages.

5. Post code that can be cut and pasted into the command line.

6. Include relevant error messages.

Hope this helps.

Greg

Answer 1

[Greg Heath]

1. There is no reason to use more than one hidden layer

2. You have created a net with 8 inputs instead of 1 8-dimensional input.

3. After creating a net view it using the command

view(net)

4. Why not just use fitnet?

help fitnet

5. After you rewrite your code you can test it on the 8-input/1-output chemical_dataset if you want to post further questions.

Hope this helps.

Thank you for formally accepting my answer

Greg

[Devarshi Rai]

</matlabcentral/answers/uploaded_files/3959/mse%20vs%20epochs.JPG> is this correct??

[Devarshi Rai]

is this correct

???

[Greg Heath]

I don't know. What data and what code?

Is your MSE normalized by the mean target variance?

[Devarshi Rai]

Input is a 78*3 matrix of 78 samples of 3 input variables Target is a 78*2 matrix of 78 samples of 2 output var. LMA.m attached is the code.

[Greg Heath]

Again, it helps if you apply your code to one of the MATLAB datasets so that results can be compared with those who take the time to help.

What version of the NNTBX are you using? Quite a bit of your code is contained in loops over matrix/vector components that MATLAB was invented to eliminate.

There is almost no use of default network properties and values.

Is this regression/curve-fitting or classification/pattern-recognition?

What happens when you apply your code to a MATLAB data set?

 help nndatasets
 [ x, t ] = iris_dataset;

[Devarshi Rai]

NNTBX ver- 6.0.4 I couldn't find any MATLAB dataset that adheres to my specifications.Input and target files are attached. It's classification- i am basically trying to determine the optimum neural network architecture{in terms of Hidden layer neurons for a 3-layered NN} by minimizing the average percentage error between predicted and target values. What do you mean by using default network properties and values? Shouldn't the neural network use these by itself?

[Greg Heath]

That doesn't make much sense to me because it is usually more important to decrease large errors than it is to decrease small errors with high relative errors.