Adding constraints to Multi-fidelity BO with discrete fidielities

[shang-zhu]

Issue description

I was following the page of Multi-fidelity BO with discrete fidelities: https://botorch.org/tutorials/discrete_multi_fidelity_bo, and try to add inequality parameter constraints to the optimizer. I only modified the optimize_acqf_mixed() and optimize_acqf() functions as below, where I tried to implement the constraint 'x[5]<0.2'. But the code ran forever and failed to output any candidates or observations. It was working well without the two constraints. Did I mess up anything here? Thank you.

Code example

def get_mfkg(model):
    curr_val_acqf = FixedFeatureAcquisitionFunction(
        acq_function=PosteriorMean(model),
        d=7,
        columns=[6],
        values=[1],
    )
    
    _, current_value = optimize_acqf(
        acq_function=curr_val_acqf,
        bounds=bounds[:,:-1],
        q=1,
        num_restarts=10 if not SMOKE_TEST else 2,
        raw_samples=1024 if not SMOKE_TEST else 4,
        options={"batch_limit": 10, "maxiter": 200},
        inequality_constraints=[(torch.tensor([5]),torch.tensor([-1]),-0.2)],
    )
        
    return qMultiFidelityKnowledgeGradient(
        model=model,
        num_fantasies=128 if not SMOKE_TEST else 2,
        current_value=current_value,
        cost_aware_utility=cost_aware_utility,
        project=project,
    )

def optimize_mfkg_and_get_observation(mfkg_acqf):
    """Optimizes MFKG and returns a new candidate, observation, and cost."""
    
    X_init = gen_one_shot_kg_initial_conditions(
        acq_function = mfkg_acqf,
        bounds=bounds,
        q=4,
        num_restarts=10,
        raw_samples=512,
    )
    candidates, _ = optimize_acqf_mixed(
        acq_function=mfkg_acqf,
        bounds=bounds,
        fixed_features_list=[{6: 0.0}, {6: 1.0}], 
        q=1,
        num_restarts=NUM_RESTARTS,
        raw_samples=RAW_SAMPLES,
        batch_initial_conditions=X_init,
        options={"batch_limit": 5, "maxiter": 200},
        inequality_constraints=[(torch.tensor([5]),torch.tensor([-1]),-0.2)],
    )
    cost = cost_model(candidates).sum()
    new_x = candidates.detach()
    new_obj = problem(new_x).unsqueeze(-1)
    print(f"candidates:\n{new_x}\n")
    print(f"observations:\n{new_obj}\n\n")
    return new_x, new_obj, cost

note that the discrete fidelities were slightly modified from the original page.

System Info

Please provide information about your setup, including

• BoTorch Version 0.6.0
• GPyTorch Version 1.6.0
• PyTorch Version 1.10.0
• mac OS Mojave Version 10.14.6

[Balandat]

Hmm since this is KG it is a high dim optimization problem so it might be that the SLSQP solver is just too slow here. But your constraint is a simple bound constraint so you can just edit the bounds you pass to the optimization instead of adding this as an inequality constraint.

[shang-zhu]

I see. I actually have a more complicated constraint: x1+x2<1, the above is more of a test for the constraint capability. Do you think it might not be feasible to optimize with KG over this constraint? If so, is there any better way to incorporate constraints for multi-fidelity BO?

[saitcakmak]

I wonder if this is related to #938. Though, as I found out in #939, that should not be an issue with qMFKG.

I think the real issue is SLSQP as @Balandat pointed out above. Here's an old comment that talks about how SLSQP tries to allocate 128gb memory when optimizing qKG. The reason for this is the time and space complexity of SLSQP is cubic in the effective dimension of the solution tensor, which in the case of qKG / qMFKG is batch_limit * (q + num_fantasies) * d. Luckily, one of these is very easy to control. I'd replace the batch_limit you pass into optimize_acqf with 1 and see if that solves the issue.