How to "retrain" Neural Networks?

Hi,

I have an algorithm to forecast time series relying on observed values of the time series in the past. The algorithm is based on several so called experts and based on their performance in the past a convex combination of the experts is used for final prediction.

In the case of Neural Networks, one of the parameters of the experts is the amount of neurons h used. Simplified I have a loop

for i=1:n

   net = newfit(inputs(:,1:i),targets(1:i),h);

   net.divideParam.trainRatio=75/100;
   net.divideParam.valRatio=25/100;
   net.trainParam.showWindow=false;
   net=train(net,inputs(:,1:i),targets(1:i));

   estimate=sim(net,x_eval);

   ...

end

So in every iteration the network trained only difers a little bit, because only one pair of (inputs,targets) is added for training, the rest remains unchanged. Is there a way to use this fact and accelerate the whole process, like "retraining"?

Thanks

Daniel

Answer 1

[Greg Heath]

Everytime you call newfit you create a new network with random initial weights. Therefore, try something like

net = newfit(inputs,targets,h);

   net.divideParam.trainRatio=75/100;

   net.divideParam.valRatio=25/100;

   net.trainParam.showWindow=false;

for i=1:n

   net = train(net,inputs(:,1:i),targets(1:i)); 

   estimate=sim(net,x_eval);  

   ...

end

Hope this helps.

Greg

[Daniel]

Hi Greg,

this should at least save some time for initiating, thank you. But the training process itself probably consumes a lot more time..

Daniel