Fixing one parameter per batch of experiments

[iandoxsee]

Hi all, I have a use case where one feature needs to be fixed among all generated experiments within a single batch.

For example, 5 experiments are generated per batch and x1 can be chosen from the range 0-100, but all 5 experiments must be executed at the same x1. In subsequent batches, a new x1 value can be used but again must be shared among all 5 experiments.

I found a reference for "process-constrained batch Bayesian optimization" and I'm wondering if this is supported with any of the existing constraints features within Botorch. Thank you for your help!

https://dl.acm.org/doi/pdf/10.5555/3294996.3295100

[Balandat]

Yeah the basic operations in that paper (i.e. fixing the value of a particular input) is supported using the fixed_features input to optimize_acqf:

botorch/botorch/optim/optimize.py

Line 429 in e68efb5

fixed_features: Optional[Dict[int, float]] = None,

Alternatively you could also use FixedFeatureAcqusitionFunction:

botorch/botorch/acquisition/fixed_feature.py

Line 52 in e68efb5

class FixedFeatureAcquisitionFunction(AcquisitionFunction):

You can use this to assemble the pc-BO(basic) algorithm from the paper after first optimizing the unconstrained acquisition function and then fixing the value for subsequent elements of the batch (just be sure to set X_pending with the previous points on the new acquisition function objects to properly account for the previous points in the batch).

However, for the pc-BO(nested) algorithm you would need to know what the proposed alpha_c^{GP-UCB} function is. I don't think this is defined anywhere in the paper... at least I haven't found it.