Sequential designs / active learning

[JannikVindeloev]

Hi @emitanaka.

cool project.
Are there current plans for including sequential designs / active learning in eddible, based on gaussian process regresion surrogate models? Also including spacefilling designs for initialization of process? What does it require to make this happen? Perhaps I could contribute?

[emitanaka]

Hi @JannikVindeloev, there is no plan for either and not sure how best edibble can accommodate those structures. Did you have thoughts of how the interface may look like?

[JannikVindeloev]

Hi @emitanaka

I had a similar idea of a framework for DoE that fitted into the tidyverse, and to not develop anything from scratch I went looking on the internet and just discovered your project yesterday.
So I'm not really familiar with your package but I think the idea is really useful. Therefore I appologize in advance if I get some your concepts or design philosophy wrong.

I'm following the guidance found here: http://jensenlab.net/automatedscience/

Planning

In model guided experiments, or adaptive learning, you would normally start with a screening design to identify active factors, i.e. factors that have an impact. This could be a Plackett–Burman Design.

When identifying important factors you would define ranges for these factors and initialize the adaptive learning cycle with a space filling design. This could be a latin hyper cube or a maximin design.

With this you would go into the lab and do your experiment and record the results.

It is not clear to me whether these design already exists or if they are easy to setup already using your framework. If not I think this is an essential addition.

Adaptive learning

In adaptive learning you would use a surrogate model for representing the real world based on the initial set of experiments.
It essential that the model can represent complex response surfaces and has a notion of uncertainty about its knowledge about the different regions in design space. The first is use to look for and exploit local / global optima, the latter is explore design space to reduce uncertainty, or some thing in between (expected improvement). Often gaussian process regression is used for this. In this case I was thinking one could add a verb in the case of gaussian process regression:

add_sequential_treatment(surrogate = "gpr", kernel = c("rbf", "matern"), noise = ..., length_scale = c(2/3, 3/5), policy = c('exploration', 'exploitation', 'expected_improvement'))

That would propose new settings of treatments based on prior knowledge (past data) that would either minimize uncertainty, look for optima, or propose treatments that would optimize expected information gains.

if you wanted more experiments in the next batch, you could use add_sequential_treatment repeatedly, changing the parameters (e.g. kernel, legnth_scale, policy)

In the multivariate case one could build a surrogate model for each response and propose new experiments based on this?

Perhaps one could also add an objective function based on the desirability metric that is used for the model guided design of experiments (https://cran.r-project.org/web/packages/desirability/vignettes/[desirability.pdf](https://cran.r-project.org/web/packages/desirability/vignettes/desirability.pdf)

It would be useful for the experimenter to be able to see the surrogate models response surface, uncertainty, and policy maps to manually add new treatments

Just my initial thoughts.

Best

Jannik