Introduction to JMP 19 new features with focus on Bayesian Optimization
Further information: https://www.jmp.com/en/software/new-release/new-in-jmp
Question 1: If we don't have historical data set, can we start a Bayesian Optimization?
Victor: Yes, you can start with Space-Filling runs to cover your experimental space
Jonas : It needs at least one row in a data set as well as some initial column information. It will add more runs to explore more of the design space in case you start with only one or a few rows.
Yes, basically the platform needs some metadata and column informations to be able to run (design role and coding at least)
Victor: I would start with a simple basic design first
Maria: Developer Tutorial on Bayesian Optimisation: https://community.jmp.com/t5/Learn-JMP-Events/Developer-Tutorial-Bayesian-Optimization/ev-p/894910?…
Victor: You can try Scoping Designs | Prism to check the variability of your response and its variability regarding the factors
Winfried: It is i the name: this is particularly useful, if you believe you can start soon to optimise. If you are still early in the process of understanding your problem, use other design tools, like screening factors or space filling designs!
Victor: Scoping Design DoE JMP Add-In - JMP User Community
Winfried: Be aware of time trend with any sequential method like BO! Remember, why we are generally use randomisation!
Victor: Also important to emphasize that DoE is not a static, one-design-answers-all-questions methodology; like Bayesian Optimization, you can progress by iterations, by augmenting your design step by step : start with a screening design, augment with higher order effects, then augment with space-filling ... The possibilities are immense, and you can combine these steps with Bayesian Optimization
Winfried: Be aware of time trend with any sequential method like BO! Remember, why we are generally use randomisation! ;)
Yes, no protection in BO like Blocking or other time-trend robust designs that we can find with DoEs !
Victor: overview of optimization methods : The Evolution of Optimization Methods | by Victor Guiller | Medium
Question 2: Is there a tool like prediction profiler which you could use for design spaces with many factors?
Victor: What is the practical difficulty behind using Prediction Profiler for 6+ factors? Optimize? You might reduce your factors space based on the importance of factors on your response(s)
Anthony: could reduce based on effect size as well. the practical significance question. if some are very low compared to the others you could reduce on those at first.
Victor: PCA might not work well using independent factors
Anthony: PCA I’ve used for reducing dimensionality in Y's. i haven't used it for multiple x's.
Victor: You can also use the Simulator, create a new dataset with predictions, and create your own visualizations.
Question 3: Including different start material product quality in the DOE designs.
You can use random blocks to allocate different batch/product qualities across your DOE runs?
Jonas: Covariate factors maybe?
Anthony: treat as a factor, covariate, blocking all come to mind as options
Manlio: you could use blocking for different materials
Victor: It depends if you want to assess the impact of batch on mean response vs. variance response (fixed vs. random effects in the model), how representative these batch qualities are, and your objectives
Question 4: How does Bayesian Optimisation handle iteration of factor addition and removal as the learning journey goes on? Design and noise factors change from experiment to experiment in my current workflow. I'm curious how BayesOpt would approach this.
Further resources:
