net.divideFcn = 'dividerand' shuffling once or in every epoch/iteration?

In the matlab NARX model, when we chose train/validation/test data randomly, does this random division happen in every epoch or just once in the beggining? I am talking about the command : net.divideFcn = 'dividerand' Thank you in advance.

Answer 1

[Mukul Joshi]

I guess just once in the beginning.

[tony gian]

so actually it is not improving the convergence/ local minima problem, it is just a measure in case your input data is structured or already classified?

[Greg Heath]

Yes. If you do not randomize the data it usually takes much longer to train because different parts of the data may have different characteristics.

An easy example is a sin wave or sawtooth.

If you actually calculate a comparison, PLEASE, PLEASE, post your results.

TYVM,

Greg

[tony gian]

Actually this should not be true in my case, because my training data do not have different characteristics, and they are not classified. But indeed dividerand works better than divideblock so i am suspecting it is shuffling after every epoch, otherwise it does not make sense.

[Greg Heath]

Wrong.

It does make sense.

Training is faster if neighboring training points are not similar.

Greg

[tony gian]

Thanks Greg, but i have one more question then. Correct me if i am wrong though, each iteration or epoch uses all the training data to determine the weights and biases. So considering our data is not initally classified or biased, why actually is training faster if neighbooring points are not similar? Shuffling each epoch makes sense because you actually use different data, but shuffling once i dont have a mathematical intuition on why it helps.Thanks in advance.

[Greg Heath]

The good thing about these differences in opinion is that they can be tested and compared.

If you do so, remember

 1. The results of regression/classification examples are 
    not necessarily valid for timeseries
 2. The results for the three types of time-series may 
    not necessarily be the same.

Hope this helps.

Greg