Issues with controlling the warm-up of BoTorch

[1kastner]

I have some constraints for my input parameters. When I use the BoTorch sampler and the first 10 samples are drawn randomly (the default setting), then afterwards it finds several solutions that satisfy my constraints. So far so good. Here an example screenshot from optuna-dashboard:

In the next step, I thought of improving the warm-up phase by copying existing trials that satisfy the constraints from another study into the BoTorch study. The samples have been generated by using a RandomSampler in combination with a dynamic search space. Here, however, in the later iterations BoTorch nearly always fails to find solutions that satisfy the constraints.

Does someone have an idea why this might be the case?

I have already screened the log files but nothing suspicious pops up.

As a sidenote and request to the developers: If more information on the BotorchSampler internals were stored in the study (SQLITE file), I assume identifying the reason could be easier. Could we store more information on the search behavior of the algorithm there?