[Feature Request] Minimization with Monte Carlo acquisition functions

[slishak-PX]

🚀 Feature Request

Analytic acquisition functions all have a maximize keyword argument to determine the direction of the optimisation problem. None of the Monte Carlo acquisition functions currently support this, so only support function maximisation.

Motivation

Is your feature request related to a problem? Please describe.
Not a problem as such, and not hard to work around, but would make a good first contribution.

Pitch

Describe the solution you'd like
Add a maximize=True keyword argument to all Monte Carlo acquisition functions.

Describe alternatives you've considered
One can simply change the sign of the function to be optimised.

Are you willing to open a pull request? (See CONTRIBUTING)
Yes, but it is not high priority so it might take some time. Logging the issue here in case another community member wishes to contribute it.

[saitcakmak]

Hi @slishak-PX. Thanks for bringing this up. Some of the analytic acquisition functions had a maximize argument from the start (as they have slightly different definitions in the literature for minimization), so we continued supporting this for the analytic acquisition functions for consistency.

Overall, I don't believe that the value add of a maximize kwarg justifies the added complexity. We have multiple options that can be used to maximize a function:

• as suggested in your post, the function can be negated
• for Gaussian posteriors, you can define a PosteriorTransform that negates the posterior mean
• you can use a LinearMCObjective with negative weights, which would simply negate the posterior samples