Parameter Constraints or post-processing of resulting candidate points?

[angyurchenko]

I use BoTorch to optimize the chemical system with multiple input parameters f(x1, x2,...xN). Due to the system's nature, I use concentrations of different units (such as %, mM, mg/mL) as the input parameters. The system has two main constraints:

1. The total volume of all components must stay below a specified limit.
2. Due to laboratory equipment limitations, we can only test specific concentration values. These values must be multiples of a minimum step size. To illustrate, If we have a chemical component A that can range from 0.01 to 1.0, and our equipment can only dispense volumes corresponding to 0.005 concentration increments, then we can only test concentrations like 0.010, 0.015, 0.020, etc. Mathematically, this means any valid concentration must satisfy: concentration % step_concentration == 0, where step_concentration is our minimum increment (0.005 in this example).

I successfully implemented the first constraint using the inequality_constraints argument of optim_acqf. However, I still cannot tackle the second constraint because, to the best of my understanding, this constraint cannot be addressed within the acquisition function optimization procedure.
To address this problem, I tried to use post-processing of the resulting concentrations by subtracting the remainder from the concentrations (concentration - concentration % step_concentration). However, this approach isn't working reliably due to floating-point precision issues in Python. It also doesn't seem feasible to convert the problem to discrete because some components can have >500 states.

Does anyone have suggestions for handling this type of discrete step constraint in BoTorch?

Thanks in advance!

[saitcakmak]

Hi @angyurchenko. BoTorch offers limited support for discrete parameters out of the box. We recently released optimize_acqf_mixed_alternating which supports optimization over integer valued discrete parameters (which will be generalized to arbitrary discrete parameters in the future).

We offer more comprehensive support for discrete parameters in Ax, via the ChoiceParameter and integer valued RangeParameter. After some transforms, these parameters are typically optimized using optimize_acqf_mixed_alternating. Under the hood, we may use continuous relaxation (if the parameter has many values), which would utilize post_processing_func to round the parameters back to the discrete values. The benefit of using Ax is that it'll automatically handle the necessary transforms for you and follow BoTorch best practices.

[angyurchenko]

Thanks for the reply! I am inclined toward post-processing since our step_concentration can vary from one experimental setup to the other. Therefore, using one "filter function" seems more straightforward, even with a potential loss in optimization accuracy due to rounding.