Auto-suggest helps you quickly narrow down your search results by suggesting possible matches as you type.

Showing results for

- JMP User Community
- :
- Discussions
- :
- Optimization with prediction profiler

- Subscribe to RSS Feed
- Mark Topic as New
- Mark Topic as Read
- Float this Topic for Current User
- Bookmark
- Subscribe
- Printer Friendly Page

- Mark as New
- Bookmark
- Subscribe
- Subscribe to RSS Feed
- Get Direct Link
- Email to a Friend
- Report Inappropriate Content

Jul 17, 2017 6:23 PM
(417 views)

Martin here.

Need some help with profiler.

It is a three factors custom design.

factor #1: Blue circle cement (0.6898~0.7431)

factor #2: Ternal white cement (0.2546~0.3009)

factor #3: sodium citrate (retarder) (0.0023~0.0093)

it is a cementitious formulation. one of the most important reponse is adhesion after water immersion. I want to optimize the adhesion, it needs to be above 1MPa.

Below is the profiler generated by JMP with maximizing desirability. adhesion after water immersion is 0.98 MPa which is not good enough to meet 1 MPa.

By manully change the propotion of three factors, profiler show below results, with adhesion after water immersion around 1.02 MPa.

So my question is what is the procedure of optimization with profiler? Do we relay on desirability function or we have to manully adjust the level of factors?

what if the desirability function fails to meet the target?

Regards

Martin

1 ACCEPTED SOLUTION

Accepted Solutions

- Mark as New
- Bookmark
- Subscribe
- Subscribe to RSS Feed
- Get Direct Link
- Email to a Friend
- Report Inappropriate Content

Jul 19, 2017 4:23 AM
(598 views)

Solution

I can tell there is a strong interaction because the effect of Sodium Citrate was null to begin with in the Prediction Profiler and then a large change in the Blue Circle level caused a large effect of Sodium Citrate. The same is true for the other factor.

Model selection is a careful process that requires all of the knowledge available. I notice from the labelling in the Parameter Estimates report that JMP is not using coded factor levels, which centers the variables. Regressing the uncoded factor levels can dramatically reduce the power of the hypothesis tests about the estimates against zero. The coding should automatically take place if you designed the experiment in JMP and retained the data table with all of the meta-data. Fit Least Squares responds to the Coding column property that was added by the JMP design platform when you clicked Make Table. Coding will help the significance of some terms. It can't hurt.

What else about model selection or model reduction puzzles you?

Learn it once, use it forever!

6 REPLIES

- Mark as New
- Bookmark
- Subscribe
- Subscribe to RSS Feed
- Get Direct Link
- Email to a Friend
- Report Inappropriate Content

Jul 18, 2017 4:11 AM
(393 views)

Are the three factors entered as mixture, continuous, or discrete numeric factors?

You appear to have a strong interaction between the blue circle and sodium citrate. That condition might lead to a local maxima. Changing the initial factor levels and repeating the optimization might lead to another maxima.

The numerical solver behind the optimization of the desirability function has settings that you can change when necessary. The settings dialog box is available through the red triangle menu.

See **Help** > **Books** > **Profilers**.

Learn it once, use it forever!

- Mark as New
- Bookmark
- Subscribe
- Subscribe to RSS Feed
- Get Direct Link
- Email to a Friend
- Report Inappropriate Content

Jul 18, 2017 4:50 PM
(379 views)

Thanks for your quick reply.

These three factors are entered as continuous factors.

How can you tell that blue circle has strong interaction with retarder sodium citrate? since all the P-value is all bigger than 0.05.

Beside the profiler optimization, the task of finding the most significant factors also puzzles me. see below JMP report.

What information I can relay on for significance of three factors and their intercation?

Thanks!

Martin

- Mark as New
- Bookmark
- Subscribe
- Subscribe to RSS Feed
- Get Direct Link
- Email to a Friend
- Report Inappropriate Content

Jul 19, 2017 4:23 AM
(599 views)

I can tell there is a strong interaction because the effect of Sodium Citrate was null to begin with in the Prediction Profiler and then a large change in the Blue Circle level caused a large effect of Sodium Citrate. The same is true for the other factor.

Model selection is a careful process that requires all of the knowledge available. I notice from the labelling in the Parameter Estimates report that JMP is not using coded factor levels, which centers the variables. Regressing the uncoded factor levels can dramatically reduce the power of the hypothesis tests about the estimates against zero. The coding should automatically take place if you designed the experiment in JMP and retained the data table with all of the meta-data. Fit Least Squares responds to the Coding column property that was added by the JMP design platform when you clicked Make Table. Coding will help the significance of some terms. It can't hurt.

What else about model selection or model reduction puzzles you?

Learn it once, use it forever!

- Mark as New
- Bookmark
- Subscribe
- Subscribe to RSS Feed
- Get Direct Link
- Email to a Friend
- Report Inappropriate Content

Jul 19, 2017 8:11 PM
(339 views)

Thanks Mark.

I wish I had your obervant eyes and quick mind to detect the problems.

Regarding** Model reduction** and **Optimization** I may have thousands of questions, could you give me a list of reading materails/books that you believe that will

improve my understanding of these two topics?

Thanks for your time

Regards

Martin

- Mark as New
- Bookmark
- Subscribe
- Subscribe to RSS Feed
- Get Direct Link
- Email to a Friend
- Report Inappropriate Content

Jul 20, 2017 6:30 AM
(306 views)

Interactions are not problems. They are opportunities! They are also part of most responses, so I am used to seeing them.

There are many sources of information about model selection and reduction. Of all the statistics topics, regression analysis has been written about often and well. In your case, let's divide the candidates into two pile: explanatory modeling and predictive modeling. Now I know that you will make predictions with your models but this term means something else. You want a good explanatory model. Furthermore, your data sets will be small and formed as a designed experiment. That narrows the choices considerably. I recommend "*Optimal Design of Experiments: A Case Study Approach*" by Perter Goos and Bradley Jones. You will discover that the design and analysis go 'hand in hand.'

And don't worry about your thousands of questions. The JMP Community is hosted on a big server...

Learn it once, use it forever!

- Mark as New
- Bookmark
- Subscribe
- Subscribe to RSS Feed
- Get Direct Link
- Email to a Friend
- Report Inappropriate Content

Jul 20, 2017 4:08 PM
(296 views)

Thanks for your advice. I'll have a read of the book you recommended.

Definitely, I will bring more questions back to JMP user community, I find it very helpful.

Regards

Martin