trainbr and data division

Hello!

I want to use trainbr algorithm (Bayesian regulation backpropagation) for some network. I haven't found information about data division for this case. Property divideParam has three standard options trainRatio, valRatio and testRatio for the default function dividerand. But the plot shows error only for test set and train set.

My question is - what is the size of data sets, that used by trainbr in fact? (dividerand default values 0.7-0.15-0.15). How my input data divided? Sorry for bad English.

Answer 1

[Greg Heath]

An ovefit network that is overtrained memorizes the training data and can perform very poorly on nontraining data. Instead of using validation set stopping to prevent poor performance on nontraining data, TRAINBR uses Bayesian Regularization. Unfortunately, TRAINBR does not allow a val set. Research has shown that in general, using both is better than using either.

So, if you wish both, use another training algorithm with the obsolete MSEREG or use the regularization option in the new MSE.

 help msereg
 help mse

Nontraining indices are calculated first. The default for TRAINBR is

 Ntst = round(0.2*N)
 Nval = 0
 Ntrn = N - Nval-Ntst

Hope this helps.

Thank you for formally accepting my answer

Greg

[Arekusandoru]

Greg, thank you for answer, but I want to know the kind of data division. For example, I have 100 samples, default division function is dividerand which default parameters are 0.7-0.15-0.15, so data usually divides in 70, 15 and 15 samples. My question is how much of this samples will be used as train set in this case (trainbr)? 70, 85 or 100?

[Greg Heath]

Repeat:

 1. Non-training data size is calculated first
 2. Unfortunately, trainbr does not allow a validation set
 3. You do the math

[Greg Heath]

The default for TRAINBR is 0.8/0.0/0.2.