Make HEBO sampler supports Define-by-Run manner, maximization, parallelization, and constant_liar

[eukaryo]

Contributor Agreements

Please read the contributor agreements and if you agree, please click the checkbox below.

• I agree to the contributor agreements.

Tip

Please follow the Quick TODO list to smoothly merge your PR.

Motivation

I would like to see a HEBO implementation that can be a drop-in replacement of Samplers in Optuna code that is written in Define-by-Run (its search space is not conditional).

I also want it to support parallelization.

Description of the changes

Modified existing HEBO package.

TODO List towards PR Merge

Please remove this section if this PR is not an addition of a new package.
Otherwise, please check the following TODO list:

• Copy ./template/ to create your package
• Replace <COPYRIGHT HOLDER> in LICENSE of your package with your name
• Fill out README.md in your package
• Add import statements of your function or class names to be used in __init__.py
• (Optional) Add from __future__ import annotations at the head of any Python files that include typing to support older Python versions
• Apply the formatter based on the tips in README.md
• Check whether your module works as intended based on the tips in README.md

[eukaryo]

#195
In the previous PR, I tried making another package, and some reviews were done.

[y0z]

• Could you provide the brief API documentation in README.md based on the latest README template?
• Making the example executable by copying and pasting it would be also helpful.

Current example (L50-L58).

search_space = {
    "x": FloatDistribution(-10, 10),
    "y": IntDistribution(0, 10),
}
sampler = HEBOSampler(search_space)
study = optuna.create_study(sampler=sampler)

Suggestion.

import optuna
import optunahub


def objective(trial: optuna.trial.Trial) -> float:
    x = trial.suggest_float("x", -10, 10)
    y = trial.suggest_int("y", -10, 10)
    return x**2 + y**2


module = optunahub.load_module("samplers/hebo")
sampler = module.HEBOSampler(search_space={
    "x": optuna.distributions.FloatDistribution(-10, 10),
    "y": optuna.distributions.IntDistribution(-10, 10),
})
study = optuna.create_study(sampler=sampler)
study.optimize(objective, n_trials=100)

print(study.best_trial.params, study.best_trial.value)

[y0z]

I've updated my comment (#196 (comment)).

[y0z]

LGTM

@nabenabe0928 If you have no further concerns, please merge this PR.

[eukaryo]

Note: I added "stateless" mode for parallelization; in the mode, it creates entire HEBO model in every trial and thus requires more computational cost. I preserved previous "define-and-run" mode for compartibility. It is not stateless but faster.

[nabenabe0928]

Thank you for the PR, I left some comments:)

[eukaryo]

@nabenabe0928 Thank you for comments! I revised the code. PTAL.

[nabenabe0928]

I left some comments, but you do not have to address all!
In the worst case, we can once merge this PR and can make another PR!

[eukaryo]

Thank you for additional comments!
I checked the current code works in the following senarios: (1)search_space is given (2) search_space=None (3) search_space=None, constant_liar=True, and n_jobs=3.
@nabenabe0928 would you please merge this PR if you have no further critical concerns?

[nabenabe0928]

LGTM, but I will check if the performance is fine just in case!

[nabenabe0928]

I will check if the shape is correct here (seems (1, N) is correct?)
Anyways, I will benchmark your code and merge this PR!

[eukaryo]

Thanks! fyi: I confirmed that simply changing line134 to hebo.observe(params, values[None, :]) or hebo.observe(params, values[None, :].T) fails.

[nabenabe0928]

I confirmed it works!