contramundum53

`scipy.stats.truncnorm.cdf` seems to be very slow, compared with `erf`. I'm not sure of the reason why `scipy.stats.truncnorm.cdf` is slow (I think it should be calculated with three calls to `erf`),...

@HideakiImamura Could you review this PR?

@ibobak Thanks for your example code! We like this feature very much. Two questions: * Do you think this can be implemented without seaborn? * Do you think the same...

This test now seems deleted. A new test with the same intention should be added.

Thanks for your suggestion! The current algorithms underlying parameter importance (fANOVA, Mean Decrease Impurity, SHAP) are inherently unstable unless the parameters are sampled uniformly (i.e. with `RandomSampler`, `QMCSampler` or `BruteForceSampler`)....

The optuna way to do that is to use conditional search space. `optuna.samplers.TPESampler` with `multivariate=True` option does handle each group of parameters in multivariate way.

As for optuna-dashboard support, there are issues like https://github.com/optuna/optuna-dashboard/issues/734.

Thanks for creating an issue! (and sorry for replying late 🙇‍♂️ ) I think this can be fixed relatively easily by modifying https://github.com/optuna/optuna/blob/master/optuna/integration/_lightgbm_tuner/optimize.py . I'm not currently having enough time...

Sorry, after a discussion, we found a potential bug (discrepancy from the original design) in the original Study class. The original intention of this issue was to use contextvars to...

In the original design, we didn't allow the `Study` class to cache trials when the user directly used `ask` and `tell`. The cache is only for `optimize` function. In that...