Auto Neural Training Process

[andreas_zaras]

Hello,

I have a data set with 4 inputs and a binary target (1-0). I use the AutoNeural tool to fit a predictive model. I set the following options:

Train Action-->Search

Number of hidden units-->1

Single Layer architecture

Final Training:Yes

Final Iterations:5

What is final training and final iterations?

The iterative process starts:

Search # 1 Single Layer Trial # 1

Training stops in iteration 8 but the model of iteration 3 is selected because of the validation misclasification error (minimum value).

A second neuron is added and a second iterative processs starts again.

Search # 2 Single Layer Trial # 1

Training stops in iteration 8 but the model of iteration 0 is selected because of the validation misclasification error (minimum value).

Then i get the following:

Final Training Training

It does 5 iterations and iteration 0 is selected.

What is that?

I guess it is related to the option mentiond before (final training:yes, final iteratons:5).

Why a third neuron is not added and the sequention increase of neuron ends at nuerons=2?

Thanks in advance,

Andreas

[DougWielenga · Accepted solution]

You can find a lot of information in the help utility for SAS Enterprise Miner.  Open the application and click on Help --> Contents and then navigate in the panel on the right to 

Node Reference

     Model 

            AutoNeural Node

and then navigate in the panel on the right to AutoNeural Node Train Properties: Increment and Search where you will read the following:

• Final Training — Use the Final Training property of the AutoNeural node to indicate whether the final model should be trained again to allow the model to converge. If the Final Training property is set to Yes, the number of iterations that are used will correspond to the value set in the Final Iterations property.
• Final Iterations — Use the Final Iterations property of the AutoNeural node to indicate the number of iterations to use when the Final Training property is set to Yes. The Final Iterations property is unavailable if the Final Training property is set to No. The Final Iterations property accepts integers greater than zero.

The information above indicates that the Final Training is only retraining the model but is not modifying the model by adding any additional nodes.  Since your Final Training requested has fewer iterations than were already fit, there is no reason to iterate since you have already found a better solution prior to Final Training.  I believe you will need to increase the number of iterations specified in Final Iterations in order to obtain some results.  

Hope this helps!

Doug