Unexpected Recommendations with Bound on a Minimization Target

[Alwalid-Abushanab]

Reproducible Example: Google Colab Notebook

I have a simple Example with 2 parameters (Name: Categorical Parameter) and (Conc: Numerical Continuous Parameter) and 1 Target (Yield: in minimization mode)

Without setting a bound on the target, the recommendations align with my expectations (Name='B' with low Conc values).
However, when I use a bound (bounds=(0, 200)), I observe unexpected recommendations where Name='A' is chosen with high Conc values, despite its Yield being high (and I am minimizing Yield).

I am able to solve this and get the expected recommendations even when bounds are defined by using objective.transform() to transform the Yield values into a normalized range before adding measurements to the campaign.

And so i have the following questions:

1. Should I always use objective.transform() before adding measurements to the campaign, especially when I have a bounded target?
2. If I use a desirability objective with multiple targets, should I call .transform() on each target separately before adding the measurements? or would calling the transform on the desirability objective be enough?

[Scienfitz]

hey, interesting observation. I didnt check out your code yet, just a few thoughts from the top of my mind when reading:

1. Could be a bug. Target.transform transforms the values into the [0,1] interval. It also mirrors the values when the target is supposed to be minimized - this might be the source of your observation. After the mirroring the values should actually be maximized, but our code keeps minimizing. I suspect we dont correctly handle this singletargetobjective-targetwithbounds situation. But the bigfix suould be a small few-liner

2. You should never have to manually use transform on your measurements. The application inverts the mirroring once again, hence I suspect why it fixes the problem. But that shouldnt be necessary in the first place from a user perspective.

3. Is there any reason you define bounds for a single target? I dont see the point for that in production runs (only makes sense to discover bugs hehe)

[Alwalid-Abushanab]

Hi, Thanks for the quick response!

I added a bound to a single target just for testing as I was having issues with the desirability objective and wanted to understand if I was doing something wrong.

But i found the issue, I was defining a bounds=(0, 2) and was getting unexpected results because the bound is not inclusive.

If I want to have a target be Less than X, would that be defined as something like bounds=(X, X.001), or is there another way?

[AdrianSosic]

Hi @Alwalid-Abushanab, thanks very much for taking the time to report this issue. I have figured out what is the source of the problem and can confirm this is indeed a bug. The reason for the behavior is a sort of "double-inversion" of the target values that happens for minimization targets as soon as a transformation is involved (which is auto-set once you define bounds). This was overlooked in #340.

The good news that the transformation part of the code will be extended anyway these days, so I think I should be able to implement a proper fix. In the meantime, we'll try to push a hotfix onto main. Let me create a standalone issue from this discussion to keep track 👍🏼 If you subscribe to it, you'll be notified as soon as the problem is resolve.

Again, thanks for your support!

[AdrianSosic]

Closed in favor of #460

[AdrianSosic]

@Alwalid-Abushanab: hotfix has been merged and is available on main. A new release will follow soon