AI Optimization in JMP Pro ByTorch

[#1]

Hi all,

Is it possible to use the JMP Pro Torch deep learning model to use it for AI Optimization for process parameter.

If it is so what are all the AI techniques are available in JMP Pro Torch deep learning.

Please suggest.

[#2]

Hi @Sankaramuthu,

Sorry, but I don't understand exactly what you would like to do :

• Do you want to use Bayesian Optimization through Python libraries for optimization ? If yes, there are several libraries available, or you can use in the meantime the Bayesian Optimization add-in developped by @yuichi_katsumur. Bayesian Optimization should be integrated in JMP 19 : DoE Bayesian optimisation
• If you want to use Deep Learning for the analysis and prediction of complex/rich data formats (text/image data), you can use Torch Deep Learning for JMP® Pro add-in developped by @russ_wolfinger. Some additional videos and tutorials are available on the Torch add-in page.

If you can be more specific about your project, type of data, objectives, context, ... it can greatly improve the quality and relevance of responses you will receive. If possible, you can also join some data with the explanations, so that people may try and recommend different approaches.

Hope this first discussion starter might help you,  

[#3]

Thanks @Victor_G for immediate response.

I have to do the optimization of process parameter for Additive manufacturing to reduce the dimensional error.

Hence looking for AI optimization option in JMP Pro Torch deep learning.

I have continuous data as factor & response.

[#4]

Ok, this very little information to help you.

Here are some questions (not an exhaustive list, please provide as much context and information as possible ! Please read  Getting correct answers to correct questions quickly) :

• How many data points, factors and responses do you have ?
• How representative are your data for the problem to solve ? How large are the ranges ? Is it only observational data (production data) ? Can you perhaps add experimental data to your production data (as production data often have narrower ranges than experimental/DOE data, that can make the optimization harder to find/realize) ?
• How are measured the responses ? How reliable/noisy are the measurement equipments ? Do you know the variability of your measurement equipment(s) ?
• What is the expected response improvement ? What is the observed response range ?
• What is your objective ? Optimization only, or also better system understanding and comparative impact evaluation of the factors ?
• What analysis have you tried already ? Why are the analysis not sufficient enough in JMP/JMP Pro ? Most of the performances problem in Machine Learning modeling are often due to data quality problems, very rarely because of having not enough model complexity. Using neural networks for tabular data is really "kill a fly with a sledgehammer" : most of the time simpler models are doing fine.
• What are your model evaluation criteria ?
• Etc...

Hope you can provide more information, 

[#5]

Hi @Sankaramuthu ,

To add to excellent replies by @Victor_G , a couple more thoughts:

1. If all of your data is tabular, XGBoost is a good potential alternative and it has a built-in hyperparameter auto-tuner.   It is available at marketplace.jmp.com along with Torch Deep Learning.

2. Torch, XGBoost, and other predictive modeling platforms have profilers that you can run from the red triangle after fitting a model and then use the Maximize Desirability functionality over your process parameters. 

3. Keep an eye out for some relevant presentations next week at JMP Discovery Summit in Berlin.