Processoptimizer

Changes

- **BREAKING**: `cook_estimator`, `has_gradients`, and `use_named_args` are moved from
`utils` to `learning`.
- **BREAKING**: `normalize_dimensions` is moved from `utils` to `space`.
- **BREAKING**: `branin`, `hart3`, `hart6`, `poly2`, and `peaks` has been changed to noisy and noiseless `ModelSystem`s.
- Implemented tests for ModelSystems.
- `gold_map` exist as `ModelSystem`.
- Sampling consolidated. There is now only one sampling function per `Dimension`.
Different sampling types (at the moment, random value sampling and Latin Hypercube
Sampling (LHS)) are handled through `Space`, referencing the sampling functions of the
`Dimesion`s
- LHS now allows for arbitrary seeding, or for random seeding, better supporting
benchmarking. The algorithm still uses a fixed seed by default.
- The module `space` uses `np.random.default_rng` as a random number generator, instead
of the deprecated `np.random.RandomState`. A bridging strategy allows it to still
accept `RandomState`s, but it will tranform them to `default_rng` for internal use.
The rest of the codebase still uses `RandomState`.
- **BREAKING**: LHS now respects priors. This means that performing LHS on a space with
a log-normal `Real` `Dimension`, or a `Categorical` `Dimension` with informative
priors will give different results in this release than in previous releases.
- **BREAKING**: The mechanism for seeding pseudorandom generators in the `space` module
have changed, meaning that, while the results are reproducible within a release, they
will not be the same as in old releases.
- Bokeh is now (again) a required installation.

Bugfixes

- Fixes to `DataDependentNoise` and `SumNoise` to avoid highly correlated
noise of the underlying noise models.
- Switched to local imports internally to avoid circular import errors.
- `NoiseModel._noise_distribution` is now a method, to allow changes of
`self._rng` to affect `self._noise_distribution` automatically.
- Removed max_features='auto' to avoid using hardcoded variables to external functions

Changes

- Added additonal model systems to the list of benchmarks and made their
structure more consistent.
- Added seeding to the noise models used for benchmarking to ensure
reproducible results when benchmarking.
- Allow addition or removal of modelled noise to the optimizer object. This
is to allow user to predict the full outcome space of a given new exp.

Bugfixes

- Fix a small number of deprecationwarnings.

Changes

- Default acquisition function changed to expected improvement (EI)
- Updated list of contributors
- Minor addition of guidance in plot_objectives()
- Implemented a major new constraint type called SumEquals. This constraint is
designed to be used for mixture experiments where a (sub)set of factors must
sum to a specific number.
- Add a module to add noise to model systems
- QoL opt.space.names added as property
- Changed default behavior of plot_objective to show uncertainty in 1D plots

Bugfixes

-

Changes

-

Bugfixes

- ParetoFront did not show full recipe for model points
- Replaced friedman_mse with squared_error

Changes

- Changed look of uncertainty-plots in plot_objective
- Added plot to only show 1d plots
- Align code in GPR module to reflect sklearn. While still supporting SKlearn
0.24.2, we have some parallel code between our local GPR and the original from
sklearn.

Bugfixes

- Model systems should now be imported as intended.

Changes

- Add Bokeh version of Pareto plot
- Make a bleeding edge installable and a stable
- Add Bokeh to list of requiered packages
- Change the call of gaussian filter for a helper illustration function
- Model systems added to help benchmark performance or teach

Bugfixes

- Remove call of plot_width and plot_height in bokeh