Add the explanation of how to sample the best expected params to FAQ #5339
Conversation
Codecov ReportAll modified and coverable lines are covered by tests ✅
Additional details and impacted files@@ Coverage Diff @@
## master #5339 +/- ##
=======================================
Coverage 89.64% 89.65%
=======================================
Files 210 210
Lines 13289 13288 -1
=======================================
Hits 11913 11913
+ Misses 1376 1375 -1 ☔ View full report in Codecov by Sentry. |
|
@contramundum53 @HideakiImamura Could you review this PR? |
| @@ -714,3 +714,66 @@ Optuna may sometimes suggest parameters evaluated in the past and if you would l | |||
|
|
|||
| study = optuna.create_study() | |||
| study.optimize(objective, n_trials=100) | |||
|
|
|||
| How can I sample the best expected parameters based on the evaluated trials? | |||
There was a problem hiding this comment.
I think some people need the parameters without sampling a new trial, so maybe it's beneficial to introduce the non-sampler way that directly uses _gp classes?
There was a problem hiding this comment.
Also it should be pointed out that this is a "hack" and it is not officially supported. It is possible that the interface changes in the future.
|
This pull request has not seen any recent activity. |
|
|
||
| As the optimization often overtunes the parameters of an objective function with an observational noise, it is convenient to sample the best parameters based on the debiased objective function. | ||
| :class:`~optuna.samplers.GPSampler` enables us to sample the best expected parameters using the mean estimation of parameters. | ||
| Note that the best expected parameters are defined as :math:`\mathrm{arg}\min_{x \in X} \mathbb{E}_{f}[f(x)]` where :math:`f` follows the Gaussian process. |
There was a problem hiding this comment.
Is
| def objective(trial, evaluate_func=True): | ||
| X = np.array([trial.suggest_float(f"x{i}", -5, 5) for i in range(2)]) | ||
| if evaluate_func: | ||
| # Could be a very expensive evaluation in practice such as a training of DNN. |
There was a problem hiding this comment.
I agree that DNN is a widely used abbreviation, but deep neural networks sounds more reader-friendly.
| # Could be a very expensive evaluation in practice such as a training of DNN. | |
| # Could be a very expensive evaluation in practice such as a training of deep neural networks. |
|
@eukaryo Could you review this PR? |
|
This pull request has not seen any recent activity. |
|
This pull request has not seen any recent activity. |
|
@nabenabe0928 Please address @nzw0301's comments. |
|
This pull request has not seen any recent activity. |
|
This pull request has not seen any recent activity. |
Motivation
As some people would like to get the best parameters that are not affected by the random noise effect, I added the guide on how to sample the expected best parameters using the
GPSamplerintroduced in v3.6.0.Description of the changes
I added the description of the way to sample the best expected parameters.