Example of incluing constraints in acquisition function acquisition.

[marcosfelt]

First, thanks for open-sourcing this project. It's amazing!

I see there's an option to pass constraints to optimize_acqf, but it's not clear to me how to get this to work in practice. Could anyone help with some examples?

Thanks!

[Balandat]

Sure - say you want to optimize a function over the unit simplex {x \in R^2: x_1>=0, x2>=0, x1+x2<=1}. Then you'd use optimize_acqf with bounds [(0, 1), (0, 1)] and inequality_constraints = [(torch.tensor([0, 1]) , torch.tensor([-1, -1]), -1)]. This translates to x[0]*(-1) + x[1]*(-1) >= -1, which if you multiply by -1 is the linear constraint x1+x2<=1 you want. You can add other linear inequality constraints by adding to the list.

You can pass equality constraints analogously.

Hope this helps!

[marcosfelt]

Okay, great. I somehow didn't see it until I wrote it down mathematically (funny ha).