[GENERAL SUPPORT]: How to run MFBO in ax?

[fightpf]

Question

Hello everyone,

I have set up the following workflow for MFBO, hoping that after loading existing data, the model will recommend a new experiment for me:

parameters.append(
    RangeParameter(
        name=fidelity_para_info["name"],
        parameter_type=ParameterType.FLOAT,
        lower=fidelity_para_info["low"],
        upper=fidelity_para_info["high"],
        is_fidelity=fidelity_para_info["is_fidelity"],
        target_value=1
    )
)
search_space = SearchSpace(parameters=parameters)
experiment = Experiment(name="my_experiment", search_space=search_space)

# (Omitting the part where old data (arms) and objective data are loaded)

botorch_kwargs = {
    "experiment": experiment,
    "data": experiment.fetch_data(),
}
model = Models.BOTORCH_MODULAR(**botorch_kwargs)
candidate = model.gen(gen_num)

However, this does not seem to work. Are there any relevant examples I can refer to?

Thanks in advance!

Please provide any relevant code snippet if applicable.

Code of Conduct

• I agree to follow this Ax's Code of Conduct

[bernardbeckerman]

Hi there,

According to the Modular BoTorchModel tutorial Appendix 2 (link), this should work as expected.

this does not seem to work

Can you copy-paste the error you're getting or what is indicating to you that this is not working as expected?

[fightpf]

Hi sir,

If I want to perform multi-objective MFBO, is it possible?

[fightpf]

I used BOTORCH_MODULAR, will get the following error message:

from botorch.models.model import Model
model = Models.BOTORCH_MODULAR(**botorch_kwargs)
candidate = model.gen(n=int(gen_num.value))

TypeError: Unexpected keyword argument target_fidelities when constructing input arguments for construct_inputs_qLogNEHVI.

When I switch to YYY, the following error appears.

from ax.modelbridge.factory import get_GPEI, get_botorch
model = get_botorch(**botorch_kwargs)
candidate = model.gen(n=int(gen_num.value))

NotImplementedError: Base BotorchModel does not support fidelity_features.

Thank you very much for your assistance.

[bernardbeckerman]

If I want to perform multi-objective MFBO, is it possible?

I'm not 100% sure, but I don't see why not. Would you please paste a minimal reproducible example so that I can recreate your setup locally and help debug?