Neural vs. Auto Neural Node in Enterprise Miner

[SlutskyFan]

In the regualar Neural Neural Network node, you can select hidden and target layer combination and activation functions ( if you select 'user' under properties-network-architecture) but it appears you can't select general architectures such as Single layer, cascade, block etc.

With the autonerual node, it appears you can select these general architectures (such as cascade or block etc. ) and you can specify activation functions (but not combination functions)

I have a couple of questions regarding these details.

1) Am I missing something? Why can't you specify (cascade, block, funnel etc.) in the regular neural net node and why can't you specify activation functions in the auto neural node?

2) With auto neural, if you select all of the activation functions that you think may apply (considering if your targe is numeric or qualitative) will EM select the optimal type of activation functions for each layer? ( is that the purpose) And again, what about combination functions?

Thanks

[David_Duling · Accepted solution]

That is basically correct. The neural node gives you properties for controlling a single hidden layer network. The autoneural node gives you an algorithm for building a multilayer network; combination function selection is not part of that algorithm. The default autoneural action is to simply train a single network to give you a baseline model. Then you can switch the train property to search to run the network search algorithm. Ultimate control is found only by using PROC NEURAL directly.

There is an almost unlimited number of possible network configurations to consider. Perhaps you can tell us what network configuration you have found useful in which situations.

Thanks!

[SlutskyFan]

This helps. I'm relatively new using EM, so I won't have anything to share for a while. One more begging question:

If you can't select the combination functions via the autoneural option, what is the default combination function for autoneural?