How can experimental designs reduce Time To Market?How can experimental designs reduce Time To Market? Thierry Taillandier & Laurent Lasbennes (Pierre Fabre)
|
Re: How can experimental designs reduce time to market?Bonjour Thierry, great work !!
I have a question: in Profiler you have Factor_4 Null which expresses no compound and the mass fraction of that compound Factor_4_W 0.125. Have you managed to find the setup that makes Profiler take into account all your constraints (e.g., if Factor_4 is Null, Factor_4_W is never >0) and not just linear mixture sum to 1?
Re: How can experimental designs reduce time to market?Hello RikkiTikkiTavi ,
Thanks for your feedback.
Good point, indeed we were unable to specify our constraints to the profiler. For each condition, we had to verify our business constraints.
We could have run a simulation to scan the entire design space and filter according to our constraints to obtain our potential hits.
I hope I have answered the question.
Re: How can experimental designs reduce time to market?Hello @RikkiTikkiTavi and @taillat ,
I see two possible options regarding this problem:
- When constructing the design : Do not include a "Null" level in the categorical factors, since this will be a situation with any compound whose mass fraction is equal to 0. Ideally, we could even add experiments to the data table to have experiments with each compound at 0, and either replicate the response value actually present in the initial design and measured for a compound with a mass fraction of 0 to all other compounds with a mass fraction of 0, or actually measure these other experimental values for compounds with a mass fraction of 0 (we are essentially adding replications since the other factors will be identical). This potentially biases the model slightly, but it allows it to have reference points for determining the trends of all compounds with a mass fraction of 0.
- During the analysis: Ajouter un script dans la table de données nommé Disallowed Combinations et écrire en JSL les combinaisons qu'on souhaite éviter. En reprenant un petit exemple avec 3 facteurs de quantité (de 0 à 1) pour 3 types de MPs (acide, base et tensioactif) et 3 facteurs de choix de MP pour chacun des types de MPs, on aurait la contrainte suivante en JSL:
Choix_acide == "Null" & Acide >= 0.01 | Base >= 0.01 & Choix_base == "Null" |
Tensioactif >= 0.01 & Choix_tensioactif == "Null";
Adding this script will influence the analysis and predictive profilers, which will recognize this constraint and avoid proposing optima for these forbidden combinations. This is the same strategy as in the discussion "How do I set up a constraint when using Bayesian Optimization?" , where simply adding a "Constraint" script to the data table limits the candidate set options of the Bayesian optimization to only the possible combinations.
Hopefully this answer will help you.
Re: How can experimental designs reduce time to market?Hello Victor, thank you for your ideas!
When using Disallowed Combinations, do I also need to remove constraints on the factors, such as the Mixture type and/or a Constraint linear constraint script? It seems that Profiler prioritizes constraints on the factors; for example, if I add both a Constraint script and a Disallowed Combinations script to the table, Profiler seems to ignore the latter.
Is there really an order of acceptance of constraints by the Profiler, or a priority rule between the different types of constraints?
Re: How can experimental designs reduce time to market?Hello @RikkiTikkiTavi ,
I need to look into this in more detail; indeed, with other linear constraints and/or mixing-type factors, it complicates the analysis and management of constraints by the Profilers.
Regarding candidate point generation, there has been an evolution in JMP's functions, and the (new) Random Linearly Constrained Uniform function allows the creation of candidate datasets with numerous constraints: linear (equalities and inequalities) and cardinality (no more than k active factors at a time in a group of n factors, for example). This presentation mentions this feature and details the JSL script (with the aforementioned function) for creating complex candidate datasets: Bayesian Optimization for Formulations Involving Complex Constraints with JMP 19 (2026-EU-30MP-2816)
However, generating the candidate point table does not create scripts that would allow restricting the modeling/analysis to the constraints used.
