Add the DDMOP (data-driven multiobjective) benchmark problems

[gialmisi]

Background

DDMOP is a suite of data-driven multiobjective optimization problems drawn from real-world applications, used widely in surrogate-assisted / offline data-driven optimization research. The problems are defined by datasets and/or surrogate models rather than clean analytic formulas. DESDEO has no implementation currently.

Heads-up (scoping): this is dataset/surrogate-driven, so it is more about combining data + surrogate evaluation into a Problem than transcribing a formula. DESDEO already has surrogate-model support (surrogatemodels/) and a simulator/external problem interface. Build on those. Partial PRs (a single DDMOP instance) are welcome.

What to do

• Obtain the DDMOP problem set (datasets / surrogate definitions) from the reference.
• For each instance, train/load the surrogate (or wire the dataset) and expose it as a Problem whose objectives query the surrogate.
• Add a constructor (e.g. ddmop(instance: int) -> Problem) registered in __init__.py.
• Add a test for at least one instance. If data is very large, discuss.

Reference

He, C., Tian, Y., Wang, H., & Jin, Y. (2020). A repository of real-world datasets for data-driven evolutionary multiobjective optimization. Complex & Intelligent Systems, 6, 189–197. (implementation https://github.com/HandingWang/DDMOP/tree/master )

Done when

• At least one DDMOP instance loads as a Problem, evaluates to objective values, and has a passing test.

[light-weaver]

Might be best to skip this one for now. The objective functions are actual blackboxes, the matlab implementation is obscured so that no one can actually see what's happening inside those files. A few ways to make things work:

1. Treat these as simulator problems (instead of surrogate problems), and call the matlab functions as blackbox simulators
2. Hunt down the actual papers where these problems originated from, and then implement those problems. Depending on the kind of problem, these can be implemented as normal desdeo problems with analytical functions or as simulator problems.

These should not be implemented as surrogate problems. That implies that a surrogate model has already been trained, which should be the job of the analyst/algorithm.