Bogdan Budescu

It's really cool that you were able to write a paper on this. Have you, by any chance, also been able to compare your method against the one employed by...

>One empirical study is in https://arxiv.org/pdf/2208.02922.pdf, in which @YIWEI-CHEN also demonstrates that when combined with early stopping, the performance can be further improved. Is this implemented in `flaml`? Perhaps it's...

On a sidenote, I can see that Luigi Nardi's [HyperMapper](https://github.com/luinardi/hypermapper) package trains a separate classification model used for anticipating whether a particular parameter configuration will yield valid results (mentioned, e.g.,...

This potential issue can be circumvented up to a certain degree using conditional sampling, but that's not very handy either. (#816, #988). E.g., I've been trying to enforce an order...

Well, it's hard for me to measure on my problem where I don't know whether or not good solutions lie on the feasibility edge, and that's why I'm asking -...

Perhaps I should try passing a custom pre-defined `ray.OptunaSearcher` instance, with the embedded space defined in terms of clean `ray.rune` APIs instead of using `flaml.tune`.

I can see this issue is turning stale. Maybe it would be an idea to make this into a feature request? I mean, the answer above clarifies the current state...

Alternatively (or, supplementary / complementary), another feature that would help solve my use case would be to find the optimal parameter configuration within a subspace defined by fixing/constraining parameters of...

Well, for my specific use case, it's not necessary to change it _during_ a single run, but, moreover, _across_ several runs. I.e., have each run explore a different subspace, but...

Any updates on this?