Is your feature request related to a problem? Please describe.
In https://github.com/CLIMADA-project/climada_python, we use Bayesian optimization for a calibration task. In our use case, we found it feasible to run the optimizer multiple times, switching between random samples and probing of the parameter space with a kappa decay. Since probing the space is relatively expensive in our case, we want to stop probing early if the optimization seems to only probe a local maximum with very little improvement.
We have been using the event subscription mechanism for that, which (in this library) was only used for logging and has been removed in v3. Now, there is no way of controlling the maximize algorithm while it runs, except manually creating random samples and probing. However, the maximize function does more and relies on private attributes for its functionality:
|
def maximize(self, init_points: int = 5, n_iter: int = 25) -> None: |
Describe the solution you'd like
Could the event subscription mechanism be re-introduced for users to control the maximize algorithm while it runs? If not, is there a recommended way of achieving it otherwise? Maybe deriving a new class from BayesianOptimization in our project?
References or alternative approaches
-/-
Additional context
We created a Controller class
https://github.com/CLIMADA-project/climada_python/blob/a423c783a6051a0d8d25be540e43f072e98bf3e1/climada/util/calibrate/bayesian_optimizer.py#L315
that is subscribed to particular events of the optimizer
https://github.com/CLIMADA-project/climada_python/blob/a423c783a6051a0d8d25be540e43f072e98bf3e1/climada/util/calibrate/bayesian_optimizer.py#L656
in addition to the logger.
Are you able and willing to implement this feature yourself and open a pull request?
Is your feature request related to a problem? Please describe.
In https://github.com/CLIMADA-project/climada_python, we use Bayesian optimization for a calibration task. In our use case, we found it feasible to run the optimizer multiple times, switching between random samples and probing of the parameter space with a kappa decay. Since probing the space is relatively expensive in our case, we want to stop probing early if the optimization seems to only probe a local maximum with very little improvement.
We have been using the event subscription mechanism for that, which (in this library) was only used for logging and has been removed in v3. Now, there is no way of controlling the maximize algorithm while it runs, except manually creating random samples and probing. However, the
maximizefunction does more and relies on private attributes for its functionality:BayesianOptimization/bayes_opt/bayesian_optimization.py
Line 348 in b2258c2
Describe the solution you'd like
Could the event subscription mechanism be re-introduced for users to control the
maximizealgorithm while it runs? If not, is there a recommended way of achieving it otherwise? Maybe deriving a new class fromBayesianOptimizationin our project?References or alternative approaches
-/-
Additional context
We created a
Controllerclasshttps://github.com/CLIMADA-project/climada_python/blob/a423c783a6051a0d8d25be540e43f072e98bf3e1/climada/util/calibrate/bayesian_optimizer.py#L315
that is subscribed to particular events of the optimizer
https://github.com/CLIMADA-project/climada_python/blob/a423c783a6051a0d8d25be540e43f072e98bf3e1/climada/util/calibrate/bayesian_optimizer.py#L656
in addition to the logger.
Are you able and willing to implement this feature yourself and open a pull request?