Showing results for 
Show  only  | Search instead for 
Did you mean: 
Choose Language Hide Translation Bar

Automatic Selection of the best of Neural Net models

What inspired this wish list request?


In JMP version 17, it is quite time consuming to randomly fit numerous neural net models to hopefully find one that fits well.  If we don't type in the seed number at the start we cannot reproduce the exact analysis later, so we cannot currently reproduce the fit versus validation plots and prediction profilers.  


What is the improvement you would like to see? 






Why is this idea important? Development of neural net models in JMP is a potentially powerful feature, but it needs improvement to make the modeler's job a bit less repetitive in searching for a neural net that well fits the data.  



Level II

For some reason it cut out the improvement I would like to see: 


1.  Please create a script that allows the user to fit a user specified number of potential neural net models.   Please have the script choose the best one based on minimizing the RMSE for a given number of factors and hidden nodes.   Script should increment the seed number and return the best fitting seed number to the user.    


2.  Please have a more advanced version of the script that allows variation in number of factors and hidden nodes.   Please provide something akin to a BIC or an AIC for scoring potential models so we can have a quantitative assessment of how many factors and how many hidden nodes are needed.  


3.  Please allow us to reproduce at any time the Neural Net model's prediction profilers, actual by predicted plots, etc..  

Super User

Great idea @GammaZulu !

Neural Networks are a complex models family, with a lot of possibilities (and complexity) and hyperparameters to tune.

In the meantime, you can also check and try the Neural Network tuning add-in from @scott_allen in order to quickly test several models in an automated way :


I hope this will help you,

Status changed to: Investigating

We are looking into this as a possible new feature in JMP.  Thanks for the suggestion!