[Feature Request] Support Exact Equality Parameter Constraints

[jkterry1]

Right now Ax only supports <= or >= as parameter constraints. Most hacks that refactor of my constraint of #153 to make it linear still requires a ==, and in general I believe such constraints are common in LPs and operations research type problems.

It's my understanding support for this is distinct and much simpler from supporting nonlinear constraints, which is why I created a separate issue.

[lena-kashtelyan]

@justinkterry, Just to make sure I understand, what would be the difference between exact equality parameter constraint and a simple fixed parameter?

[jkterry1]

@lena-kashtelyan Most of the use cases I've seen involving == look something like param_a*parma_b == constant,

[jkterry1]

Or in the linear case you could have aparam_a+bparam_b == constant. These things all come up in cases where different parameters values consume different portions of a finite assets.

[Balandat]

This one is much less challenging to support than general nonlinear constraints since a linear constraint essentially means intersecting the constraint polytope with a lower-dimensional subspace, so the resulting relative interior of the intersection is still a convex polytope in that subspace.

We have some of the necessary components for doing this implemented on the BoTorch end (namely sampling from such spaces). Methodologically there shouldn't be anything else overly challenging, but I suspect it might require quite a bit of engineering work to hook this up properly.

[rzc331]

@lena-kashtelyan Most of the use cases I've seen involving == look something like param_a*parma_b == constant,

I want to second it, I tried using an alternative way:

con_1 = SumConstraint(parameters=[salt_1, salt_2, additive_1, additive_2, solvent_1, solvent_2],
is_upper_bound=True, bound=1)

con_2 = SumConstraint(parameters=[salt_1, salt_2, additive_1, additive_2, solvent_1, solvent_2],
is_upper_bound=False, bound=1)

Theoretically, it should work since the inequality is not rigorous. But I got no luck:

raise SearchSpaceExhausted(
ax.exceptions.core.SearchSpaceExhausted: Rejection sampling error (specified maximum draws (10000) exhausted, without finding sufficiently many (1) candidates). This likely means that there are no new points left in the search space.

I interprete it as that the points satisfying the equality are very sparse.

It would be fantastic if the equality constraint could be implemented officially.

Thanks in advance!

[Balandat]

I interprete it as that the points satisfying the equality are very sparse.

Yes, this currently uses rejection sampling by default so with an equality constraint the feasible set has zero volume and this will be the outcome.

We have added some features to BoTorch under the hood that allows sampling in different ways and we plan to hook things up in Ax to allow parameter constraints. That will take a bit of work though since it touches storage layers and a lot of other places.

[lena-kashtelyan]

We will now be tracking wishlist items / feature requests in a master issue for improved visibility: #566. Of course please feel free to still open new feature requests issues; we'll take care of thinking them through and adding them to the master issue.