Parameter constraints

[nikitrian]

Hello,

I am adding parameter constraints to my problem as inequality_constraints and I am using ConstrainedExpectedImprovement as an acquisition function, in order to do that. I was wondering if I can use inequality_constraints with other acquisition functions such as UCB, MaxValueEntropy and Probability of Improvement.

Thank you!

[saitcakmak]

Hi @nikitrian. The constraint in ConstrainedEI refers to the outcome constraints. Parameter constraints only affect how you optimize the acquisition function (i.e., how optimize_acqf functions), so the choice of the acquisition function should not matter.

[nikitrian]

I see.. but when I do that the parameter constraints are violated
inequality_constraints = ["(x[..., 1]-x[..., 0])"]

acquisition = UpperConfidenceBound(gp, inequality_constraints=inequality_constraints, beta=15, maximize=False)

I want x0<=x1

[saitcakmak]

Parameter constraints should not be specified while constructing the acquisition function. You need to pass them to optimize_acqf using the syntax described in https://github.com/pytorch/botorch/blob/main/botorch/optim/optimize.py#L451-L456

If your goal is to optimize a given objective with several different acquisition functions, I'd highly recommend using Ax, which has a very convenient API for specifying constraints: https://ax.dev/tutorials/gpei_hartmann_service.html

We also have a tutorial showing how to use a custom BoTorch acquisition function with Ax: https://botorch.org/tutorials/custom_acquisition

[nikitrian]

Very helpful, thank you!