topic Optimal Designs in Discussions

Optimal Designs

Thommy7571 — Thu, 06 Jun 2024 16:20:55 GMT

Hallo, kann JMP auch optimale Designs erstellen und auswerten? D-, E-, I-Designs usw.?

Welches optimale Design ist als Screening Design brauchbar?

Welche Anzahl von Läufen ist sinnvoll?

Kann dieses erweitert und dann für eine komplexere Auswertung verwendet werden?

Was für ein optimales Design ist für die komplexere Auswertung zu empfehlen?

Welche Anzahl von Läufen ist dann sinnvoll?

Danke.

Thommy7571

Re: Optimale Designs

Victor_G — Thu, 06 Jun 2024 14:56:09 GMT

Hi @Thommy7571,

Optimal designs are available in JMP under the Custom design platform.

By default, JMP will create a design with a recommended optimality criterion based on the model assumed : Optimality Criteria (jmp.com) But you can change the optimality criterion manually by clicking in the red triangle next to Custom design, go to "Optimality Criterion", and from there select the one you want (between D-, I-, A- and Alias-optimality criteria).
For screening designs, D- and A- optimal designs are recommended. Bradley Jones recommends using A-Optimal designs for screening designs, as they offer similar performances to D-Optimal designs with higher flexibility if you want to put more emphasis on certain terms/effects estimation (for example main effects vs. interactions). See his presentation here : 21st Century Screening Designs (2020-US-45MP-538) - JMP User Community
There is no perfect situation concerning the number of runs, it's a compromise between your available experimental budget/ressources, your objectives and the confidence/precision you want in the results. One good option is to create several designs, and compare them using the platform Compare Designs (jmp.com). You can also generate several designs and compare them quickly (to assess the benefits of increasing the number of runs for example) using the Design Explorer (jmp.com).
One design may not answer all your questions/needs.
Think sequentially and iterate : you can start with a screening design (Optimal design, Definitive Screening design if applicable, others, ...), then augment it to uncover/detect higher order effects through an optimisation design (response surface type), and if you have non-linear/complex relationships between factors, you can also augment your design through a Space-Filling approach, to benefit from various Machine Learning methods in the analysis of results.

If you would like to have a follow-up with more precise questions on a specific topic, feel free to answer :)

Hope this first answer will help you,

Re: Optimale Designs

Thommy7571 — Thu, 13 Jun 2024 12:14:26 GMT

Hallo,

Kann ein Definitive Screening Design (DSD) in ein A-optimales design umgewandelt werden? Falls ja, kann das sinnvoll sein?

Ich habe versucht einem DSD faktorielle Punkte hinzuzufügen, aber bei der Umwandlung in ein optimales Design wurden die Läufe mit Null wieder entfernt.

Grüße

Thommy7571

Re: Optimale Designs

Victor_G — Thu, 13 Jun 2024 12:57:35 GMT

Hi @Thommy7571,

The construction of Definitive Screening Designs and Optimal designs are very different :

The construction of Definitive Screening Design relies on the use of pre-determined conference matrices (Xiao et al., 2012) : Structure of Definitive Screening Designs
The construction of Optimal design relies on algorithmic iterative refinement : starting from a set of random points in the defined experimental space, the coordinate-exchange algorithm replace some factors values of the points in order to optimize the optimality criterion chosen : Coordinate-Exchange Algorithm
And a visualization/complementary explanation here : https://www.linkedin.com/posts/victorguiller_designofexperiments-optimaldesigns-activity-7163071856332234752-z0wb

So I'm not sure what you imply with "converting" a design type into another, but you can augment a DSD (or any other design) into an A-Optimal design.
Once again, the augmentation choices are different and rely on different techniques :

Algorithmic computations for "Space-Filling" and "Augment" (optimal design augmentation) types
Pre-determined and specific points for "Replicate", "Add Centerpoints", "Fold Over" and "Add Axial",

so you can indeed expect different augmented points proposals betwwen these augmentation techniques.

Hope this answer will help you,