Nice work on NEXTorch! Seems like a compelling idea, and some difficulty with implementing human-in-the-loop optimization via Ax is what led me here.
Something very important for my use-case is inequality constraints for composition-based Bayesian optimization (also prevalence-based, fractional-based, etc., i.e. everything needs to sum to 1). For example:
A + B + C <= 1
where A, B, and C are parameters to be optimized.
For me, there would technically also be D, the final parameter is determined automatically outside of the optimization loop via:
D = 1 - sum([A, B, C])
In other words, safe to ignore D, but relevant for the context of a composition-based optimization.
Is it possible to specify an inequality constraint within the current implementation of NEXTorch?
Nice work on NEXTorch! Seems like a compelling idea, and some difficulty with implementing human-in-the-loop optimization via Ax is what led me here.
Something very important for my use-case is inequality constraints for composition-based Bayesian optimization (also prevalence-based, fractional-based, etc., i.e. everything needs to sum to
1
). For example:where
A
,B
, andC
are parameters to be optimized.For me, there would technically also be
D
, the final parameter is determined automatically outside of the optimization loop via:In other words, safe to ignore
D
, but relevant for the context of a composition-based optimization.Is it possible to specify an inequality constraint within the current implementation of NEXTorch?