BookmarkSubscribeRSS Feed
sharonxi1205

Community Member

Joined:

Dec 1, 2018

what's differnence of "restart" and "rerun" when using neutural network to predic

I have encounter a problem, i am using "Nuetural Network" to do prediction and trying different combination of activation function. I have set the randomseed=7. so that every restart of the model generate the same result as following.Tan10G10Tan10G10.JPG

 

 

 

But if it is not restart but rerun in the same window using "rerun" not "restart", then the result is different every time. so i wonder what cause the random result.?Tan10G10Tan10G10_Rerun.JPG

 

 

 

1 REPLY
Highlighted
phil_kay

Staff

Joined:

Jul 22, 2014

Re: what's differnence of "restart" and "rerun" when using neutural network to p

Okay, so just to clarify this is an example with Big Class:

Open( "$SAMPLE_DATA/Big Class.jmp" );
Neural(
	Y( :weight ),
	X( :age, :sex, :height ),
	Informative Missing( 0 ),
	Validation Method( "Holdback", 0.3333 ),
	Set Random Seed( 7 ),
	Fit( NTanH( 3 ) ),
	Fit( NTanH( 3 ) )
);

This is what you get if you repeat "Go" from the neural net model launch dialog. We have set a random seed of 7 but the two models are different. If we "redo" this analysis we get exactly the same result for the two models.

Yes, I can see that you would not expect that.

With neural nets there is a random nature to the fitting. Setting the random seed ensures that this random behaviour can be reproduced. However, there is still a randomisation between instances of the model launch. You can see that this randomisation is also reproduced by setting the random seed.

Is this is a problem. Do you need repeat instances of the model fit in the same dialog to be exactly the same. I can't imagine why that would be helpful?

Regards,

Phil