how to test and improve multi layer perceptron

My project is about apnea detection based on ECG features; I have 11 features for each ECG and my data set is 100 Apnea and 100 normal signal. I choosed newff to implement the classifier with one hidden layer of 8 neurons and I have divided the data into training and testing. The problem is that I assumed '1' as apnea target and '0' as normal target but when it comes to testing the output is not convenient at all Here is my code where R is the training set,T is its target and S is the testing set

function[Output,trained_net,stats]= net_train(R,T,S)

net = newff(R', T, [8], {'tansig' 'logsig'}, 'traingd', ... '', 'mse', {}, {}, '');

net=init(net);

net.trainparam.min_grad=0;

net.trainparam.epochs=1000;

[trained_net,stats]=train(net,R',T);

Output=sim(trained_net,S');

end

[Greg Heath]

Please explain why the {0,1} target is giving you problems.

It shouldn't.

[amanda hachem]

I improved my algorithm and when training I reached 95% accuracy but the testing is still not working (almost 35%).

function[Y,net]=MLPNN(R,T,S)

rand('seed', 672880951)

net = newff(minmax( R ),[11,11,1],{'tansig','logsig','logsig'});

net.trainFcn='trainscg';

net.trainParam.epochs=1000;

net=train(net,R,T);

Y=round(sim(net,S));

end

R the training data is composed of 50 consecutive apnea records and 50 consecutive normal records and so does the testing data S.

the output should be like [ones(1,50) zeros(1,50)] but it does not. Any idea?

Answer 1

[Greg Heath]

NEWPR is the version of NEWFF that should be used for classification/pattern-recognition

Remove ending semicolons to see command line printout:

 target           = ind2vec(trueclassindices);% Use for training
trueclassindices = vec2ind(target);
...
output = net(input);
predictedclassindices = vec2ind(output);
err = predictedclassindices ~= trueclassindices; % ones and zeros
Nerr = sum(err);
PctErr = 100*Nerr/N

classification breakdowns of the trn/val/tst subsets can be obtained using the training record tr obtained from

 [ net tr output error  ] = train(net,input,target);
 % output = net(input);
 % error = target-output;

 tr = tr    % No semicolon to see the goodies

Hope this helps.

Thank you for formally accepting my answer

Greg

[amanda hachem]

thanks for helping it worked but i have another question.

Now i'm using newrb to classify using Radial basis but it immediately reaches minimum gradient. What do i have to change?

Thank you for your help