Adding Desirability functions as objectives

[LukasHebing]

• added a file with data-models 3 different Desirability functions: Increasing, Decreasing, and Peak Desirability
• Some options for the function design (bounds, clipping, concave/convex form
• Added torch executables into torch_tools.py
• Tests for data-models and objectives added
• A demonstration notebook, on how the functions look is added as example

[LukasHebing]

There are some Linting errors. I dont know why pyright doesnt find anything for all the objective classes? Even the old ones:

Error: Cannot access member "peak_position" for type "MinimizeSigmoidObjective"
Member "peak_position" is unknown (reportGeneralTypeIssues)
/home/runner/work/bofire/bofire/bofire/utils/torch_tools.py:494:70 - error: Cannot access member "peak_position" for type "TargetObjective"
Member "peak_position" is unknown (reportGeneralTypeIssues)

[LukasHebing]

entmoot on python 3.9 is still responsible for the failing tests (see #493)

[LukasHebing]

@jduerholt: I really don`t know, how to fix the linting errors of the data-models. Field definitions of the other objectives are unchanged, still pyright wants to validate fields from other objectives, e.g.:

Error: Cannot access member "lower_bound" for type "CloseToTargetObjective"
Member "lower_bound" is unknown (reportGeneralTypeIssues)
/home/runner/work/bofire/bofire/bofire/utils/torch_tools.py:503:58 - error: Cannot access member "lower_bound" for type "ConstrainedCategoricalObjective"
Member "lower_bound" is unknown (reportGeneralTypeIssues)

[jduerholt]

Hi @LukasHebing, I had a day off yesterday, I will have a look today. Best, Johannes

[jduerholt]

Hi @LukasHebing, I had a day off yesterday, I will have a look today. Best, Johannes PS: We merged the PR that drops the python 3.9 support. So the errors due to entmoot should be gone, if you merge main in.

[jduerholt]

@LukasHebing: Thanks!

I did a first pass with some general remarks. Regarding the linting: Maybe it stems from the not provided type argument for the new objectives?

[LukasHebing]

The failing test is due to changes in the main branch:
features_to_idx_mapper in map_HammingDistanceKernel seems to cause a problem

[Edit]
That failed test disappeared. Very strange

[LukasHebing]

@jduerholt:
Do you have any idea, why tests in test_ask.py are failing, which have nothing to do with this PR? It seems to originate in the acquisition function qNegInt. But specifications of these tests is very difficult. I think, this is not due to any of this changes.

_________ test_ask_single_objective[SoboStrategy-spec3-False-False-2] __________

cls = <class 'bofire.data_models.strategies.predictives.sobo.SoboStrategy'>
spec = {'acquisition_function': qNegIntPosVar(type='qNegIntPosVar', n_mc_samples=128, weights={'y_1': 0.5, 'y_2': 0.5}), 'cat...ctive(type='MaximizeObjective', w=1.0, bounds=[0, 1]))]), constraints=Constraints(type='Constraints', constraints=[]))}
categorical = False, descriptor = False, candidate_count = 2

@pytest.mark.parametrize(
    "cls, spec, categorical, descriptor, candidate_count",
    [
        (cls, specs, categorical, descriptor, random.randint(1, 2))
        for cls, specs in STRATEGY_SPECS_SINGLE_OBJECTIVE.items()
        for categorical in [True, False]
        for descriptor in [True, False]
    ],
)
# @pytest.mark.slow
def test_ask_single_objective(cls, spec, categorical, descriptor, candidate_count):
    # generate data
    benchmark = Ackley(categorical=categorical, descriptor=descriptor)
    random_strategy = RandomStrategy(
        data_model=RandomStrategyDataModel(domain=benchmark.domain),
    )
    experiments = benchmark.f(random_strategy.ask(10), return_complete=True)

    # set up of the strategy

  data_model = cls(**{**spec, "domain": benchmark.domain})

E pydantic_core._pydantic_core.ValidationError: 7 validation errors for SoboStrategy
E acquisition_function.qNEI
E Input should be a valid dictionary or instance of qNEI [type=model_type, input_value=qNegIntPosVar(type='qNegI...'y_1': 0.5, 'y_2': 0.5}), input_type=qNegIntPosVar]
E For further information visit https://errors.pydantic.dev/2.10/v/model_type
E acquisition_function.qEI
E Input should be a valid dictionary or instance of qEI [type=model_type, input_value=qNegIntPosVar(type='qNegI...'y_1': 0.5, 'y_2': 0.5}), input_type=qNegIntPosVar]
E For further information visit https://errors.pydantic.dev/2.10/v/model_type
E acquisition_function.qSR
E Input should be a valid dictionary or instance of qSR [type=model_type, input_value=qNegIntPosVar(type='qNegI...'y_1': 0.5, 'y_2': 0.5}), input_type=qNegIntPosVar]
E For further information visit https://errors.pydantic.dev/2.10/v/model_type
E acquisition_function.qUCB
E Input should be a valid dictionary or instance of qUCB [type=model_type, input_value=qNegIntPosVar(type='qNegI...'y_1': 0.5, 'y_2': 0.5}), input_type=qNegIntPosVar]
E For further information visit https://errors.pydantic.dev/2.10/v/model_type
E acquisition_function.qPI
E Input should be a valid dictionary or instance of qPI [type=model_type, input_value=qNegIntPosVar(type='qNegI...'y_1': 0.5, 'y_2': 0.5}), input_type=qNegIntPosVar]
E For further information visit https://errors.pydantic.dev/2.10/v/model_type
E acquisition_function.qLogEI
E Input should be a valid dictionary or instance of qLogEI [type=model_type, input_value=qNegIntPosVar(type='qNegI...'y_1': 0.5, 'y_2': 0.5}), input_type=qNegIntPosVar]
E For further information visit https://errors.pydantic.dev/2.10/v/model_type
E acquisition_function.qLogNEI
E Input should be a valid dictionary or instance of qLogNEI [type=model_type, input_value=qNegIntPosVar(type='qNegI...'y_1': 0.5, 'y_2': 0.5}), input_type=qNegIntPosVar]
E For further information visit https://errors.pydantic.dev/2.10/v/model_type

[jduerholt]

This is kind of a Heisenbug, I have to take some time and tidy up these tests, it some kind of legacy back from everest and is very messy. Just ignore them for now.

[jduerholt]

That failed test disappeared. Very strange

It has been fixed on main, I assume it is vanished due to this ;)

[LukasHebing]

Works now, only the linting error remains.
@jduerholt: Can this be skipped?

[jduerholt]

Hi Lukas,

looks good to me. Only some smaller comments regarding inheritance and testing.

Best,

Johannes

[LukasHebing]

@jduerholt : If we ignore the linting (don`t know how to fix that), I fixed everything. Let me know, if we can merge

[jduerholt]

Looks good to me. Thanks!