rbnn functions works with the small data set but the one i want to use doesnot work at all. the simulation is too slow and i have to resatart my matlab again. need help what should i do

%reading train and test data from seperate text files containing integers 

P=train(:,1:7); %13000*8
T=train(:,8); % 13000*1

a=test(:,1:7); %5545*7
s=test(:,8); %5545*1

[pn,minp,maxp,tn,mint,maxt]=premnmx(P',T');
[an,mina,maxa,sn,mins,maxs]=premnmx(a',s');

net = newrb(pn,tn)

t = sim(net,an);  
X=postmnmx(t',mins,maxs)
error = X - s;
[s X];
%subplot(1,2,1)
plot(1:length(a),s,'.-r',1:length(a),X,'.-b')
% subplot(1,2,2)
% plot(error)

Problem 2: When I use newgrnn it works fine although with huge errors. how to add more hidden layer in newgrnn. if there is an inbuild command for it will be very helpful so far i found out that the matlab grnn network has only 1 default hidden layer. Thanks

Answer 1

[Greg Heath]

 1. You probably do not need more than 10 to 30 times as many input/target training pairs as you have weights and parameters to estimate. Clustering might help reduce the size of the training set.

 2. You might also think of designing an ensemble of several nets designed on fractions of the data. Then combine the outputs (e.g., average or a linear oombination)

 3. Typo   13000*8

 4.Errors   subplot(1,2,1), subplot(1,2,2)

 5. I prefer standardization ( zscore or mapstd) for bad data detection

 6. Delete or modify bad or exteme data beyond ~ mean +/- 3*stdv

 7. You can use radbas functions in fitnet/newfit, patternnet/newpr, or feedforwardnet/newff if you want more hidden layers.

Hope this helps.

Thank you for formally accepting my answer

Greg

[John Albert]

Thank you