MATLAB: How to generate initial weights for back-propagation neural network (BPNN) using MATLAB R2012a

bpnnDeep Learning ToolboxMATLABweight

I am trying to predict future values using BPNN, with 1 input and output neuron respectively and also 1 layer of hidden layer. And I change the number of hidden neurons from 2 to 10. For example, I set the number of hidden neurons (N) as 10 first. So, I want to generate the weights (10 weights from input to hidden, and 10 from hidden to output),and I write this:
s=rng; m=rand(10,2)
Is this the correct way? However, when I change the N to 7, it gives me the same predicted values as N=10.

Best Answer

The reason why the answers are different is because the state of the random number generator changes each time it is called.
Therefore if you want reproducibility, initialize the state of the rng before the first call of configure or train.
Most of my serious designs involve searching over multiple designs for the best combination of random initial weights, random trn/val/tst data divisions and number of hidden nodes.
I have posted scores of examples in the NEWSGROUP and ANSWERS. Search on
 greg rng('default')
 or
 greg rng(0)
Hope this helps
Thank you for formally accepting my answer
Greg
Related Question