Baybe

Added
- Continuous linear constraints have been consolidated in the new
`ContinuousLinearConstraint` class

Changed
- `get_surrogate` now also returns the model for transformed single targets or
desirability objectives

Fixed
- Unsafe name-based matching of columns in `get_comp_rep_parameter_indices`

Deprecations
- `ContinuousLinearEqualityConstraint` and `ContinuousLinearInequalityConstraint`
replaced by `ContinuousLinearConstraint` with the corresponding `operator` keyword

Breaking Changes
- The public methods of `Surrogate` models now operate on dataframes in experimental
representation instead of tensors in computational representation
- `Surrogate.posterior` models now returns a `Posterior` object
- `param_bounds_comp` of `SearchSpace`, `SubspaceDiscrete` and `SubspaceContinuous` has
been replaced with `comp_rep_bounds`, which returns a dataframe

Added
- `py.typed` file to enable the use of type checkers on the user side
- `IndependentGaussianSurrogate` base class for surrogate models providing independent
Gaussian posteriors for all candidates (cannot be used for batch prediction)
- `comp_rep_columns` property for `Parameter`, `SearchSpace`, `SubspaceDiscrete`
and `SubspaceContinuous` classes
- New mechanisms for surrogate input/output scaling configurable per class
- `SurrogateProtocol` as an interface for user-defined surrogate architectures
- Support for binary targets via `BinaryTarget` class
- Support for bandit optimization via `BetaBernoulliMultiArmedBanditSurrogate` class
- Bandit optimization example
- `qThompsonSampling` acquisition function
- `BetaPrior` class
- `recommend` now accepts the `pending_experiments` argument, informing the algorithm
about points that were already selected for evaluation
- Pure recommenders now have the `allow_recommending_pending_experiments` flag,
controlling whether pending experiments are excluded from candidates in purely
discrete search spaces
- `get_surrogate` and `posterior` methods to `Campaign`
- `tenacity` test dependency
- Multi-version documentation

Changed
- The transition from experimental to computational representation no longer happens
in the recommender but in the surrogate
- Fallback models created by `catch_constant_targets` are stored outside the surrogate
- `to_tensor` now also handles `numpy` arrays
- `MIN` mode of `NumericalTarget` is now implemented via the acquisition function
instead of negating the computational representation
- Search spaces now store their parameters in alphabetical order by name
- Improvement-based acquisition functions now consider the maximum posterior mean
instead of the maximum noisy measurement as reference value
- Iteration tests now attempt up to 5 repeated executions if they fail due to numerical
reasons

Fixed
- `CategoricalParameter` and `TaskParameter` no longer incorrectly coerce a single
string input to categories/tasks
- `farthest_point_sampling` no longer depends on the provided point order
- Batch predictions for `RandomForestSurrogate`
- Surrogates providing only marginal posterior information can no longer be used for
batch recommendation
- `SearchSpace.from_dataframe` now creates a proper empty discrete subspace without
index when called with continuous parameters only
- Metadata updates are now only triggered when a discrete subspace is present
- Unintended reordering of discrete search space parts for recommendations obtained
with `BotorchRecommender`

Removed
- `register_custom_architecture` decorator
- `Scalar` and `DefaultScaler` classes

Deprecations
- The role of `register_custom_architecture` has been taken over by
`baybe.surrogates.base.SurrogateProtocol`
- `BayesianRecommender.surrogate_model` has been replaced with `get_surrogate`

Breaking Changes
- Providing an explicit `batch_size` is now mandatory when asking for recommendations
- `RecommenderProtocol.recommend` now accepts an optional `Objective`
- `RecommenderProtocol.recommend` now expects training data to be provided as a single
dataframe in experimental representation instead of two separate dataframes in
computational representation
- `Parameter.is_numeric` has been replaced with `Parameter.is_numerical`
- `DiscreteParameter.transform_rep_exp2comp` has been replaced with
`DiscreteParameter.transform`
- `filter_attributes` has been replaced with `match_attributes`

Added
- `Surrogate` base class now exposes a `to_botorch` method
- `SubspaceDiscrete.to_searchspace` and `SubspaceContinuous.to_searchspace`
convenience constructor
- Validators for `Campaign` attributes
- `_optional` subpackage for managing optional dependencies
- New acquisition functions for active learning: `qNIPV` (negative integrated posterior
variance) and `PSTD` (posterior standard deviation)
- Acquisition function: `qKG` (knowledge gradient)
- Abstract `ContinuousNonlinearConstraint` class
- Abstract `CardinalityConstraint` class and
`DiscreteCardinalityConstraint`/`ContinuousCardinalityConstraint` subclasses
- Uniform sampling mechanism for continuous spaces with cardinality constraints
- `register_hooks` utility enabling user-defined augmentation of arbitrary callables
- `transform` methods of `SearchSpace`, `SubspaceDiscrete` and `SubspaceContinuous`
now take additional `allow_missing` and `allow_extra` keyword arguments
- More details to the transfer learning user guide
- Activated doctests
- `SubspaceDiscrete.from_parameter`, `SubspaceContinuous.from_parameter`,
`SubspaceContinuous.from_product` and `SearchSpace.from_parameter`
convenience constructors
- `DiscreteParameter.to_subspace`, `ContinuousParameter.to_subspace` and
`Parameter.to_searchspace` convenience constructors
- Utilities for permutation and dependency data augmentation
- Validation and translation tests for kernels
- `BasicKernel` and `CompositeKernel` base classes
- Activated `pre-commit.ci` with auto-update
- User guide for active learning
- Polars expressions for `DiscreteSumConstraint`, `DiscreteProductConstraint`,
`DiscreteExcludeConstraint`, `DiscreteLinkedParametersConstraint` and
`DiscreteNoLabelDuplicatesConstraint`
- Discrete search space Cartesian product can be created lazily via Polars
- Examples demonstrating the `register_hooks` utility: basic registration mechanism,
monitoring the probability of improvement, and automatic campaign stopping
- Documentation building now uses a lockfile to fix the exact environment

Changed
- Passing an `Objective` to `Campaign` is now optional
- `GaussianProcessSurrogate` models are no longer wrapped when cast to BoTorch
- Restrict upper versions of main dependencies, motivated by major `numpy` release
- Sampling methods in `qNIPV` and `BotorchRecommender` are now specified via
`DiscreteSamplingMethod` enum
- `Interval` class now supports degenerate intervals containing only one element
- `add_fake_results` now directly processes `Target` objects instead of a `Campaign`
- `path` argument in plotting utility is now optional and defaults to `Path(".")`
- `UnusedObjectWarning` by non-predictive recommenders is now ignored during simulations
- The default kernel factory now avoids strong jumps by linearly interpolating between
two fixed low and high dimensional prior regimes
- The previous default kernel factory has been renamed to `EDBOKernelFactory` and now
fully reflects the original logic
- The default acquisition function has been changed from `qEI` to `qLogEI` for improved
numerical stability

Removed
- Support for Python 3.9 removed due to new [BoTorch requirements](https://github.com/pytorch/botorch/pull/2293)
and guidelines from [Scientific Python](https://scientific-python.org/specs/spec-0000/)
- Linter `typos` for spellchecking

Fixed
- `sequential` flag of `SequentialGreedyRecommender` is now set to `True`
- Serialization bug related to class layout of `SKLearnClusteringRecommender`
- `MetaRecommender`s no longer trigger warnings about non-empty objectives or
measurements when calling a `NonPredictiveRecommender`
- Bug introduced in 0.9.0 (PR 221, commit 3078f3), where arguments to `to_gpytorch`
are not passed on to the GPyTorch kernels
- Positive-valued kernel attributes are now correctly handled by validators
and hypothesis strategies
- As a temporary workaround to compensate for missing `IndexKernel` priors,
`fit_gpytorch_mll_torch` is used instead of `fit_gpytorch_mll` when a `TaskParameter`
is present, which acts as regularization via early stopping during model fitting

Deprecations
- `SequentialGreedyRecommender` class replaced with `BotorchRecommender`
- `SubspaceContinuous.samples_random` has been replaced with
`SubspaceContinuous.sample_uniform`
- `SubspaceContinuous.samples_full_factorial` has been replaced with
`SubspaceContinuous.sample_from_full_factorial`
- Passing a dataframe via the `data` argument to the `transform` methods of
`SearchSpace`, `SubspaceDiscrete` and `SubspaceContinuous` is no longer possible.
The dataframe must now be passed as positional argument.
- The new `allow_extra` flag is automatically set to `True` in `transform` methods
of search space classes when left unspecified

Expired Deprecations (from 0.7.*)
- `Interval.is_finite` property
- Specifying target configs without type information
- Specifying parameters/constraints at the top level of a campaign configs
- Passing `numerical_measurements_must_be_within_tolerance` to `Campaign`
- `batch_quantity` argument
- Passing `allow_repeated_recommendations` or `allow_recommending_already_measured`
to `MetaRecommender` (or former `Strategy`)
- `*Strategy` classes and `baybe.strategies` subpackage
- Specifying `MetaRecommender` (or former `Strategy`) configs without type information

Changed
- Discrete searchspace memory estimate is now natively represented in bytes

Fixed
- Non-GP surrogates not working with `deepcopy` and the simulation package due to
slotted base class
- Datatype inconsistencies for various parameters' `values` and `comp_df` and
`SubSelectionCondition`'s `selection` related to floating point precision

Added
- Class hierarchy for objectives
- `AdditiveKernel`, `LinearKernel`, `MaternKernel`, `PeriodicKernel`,
`PiecewisePolynomialKernel`, `PolynomialKernel`, `ProductKernel`, `RBFKernel`,
`RFFKernel`, `RQKernel`, `ScaleKernel` classes
- `KernelFactory` protocol enabling context-dependent construction of kernels
- Preset mechanism for `GaussianProcessSurrogate`
- `hypothesis` strategies and roundtrip test for kernels, constraints, objectives,
priors and acquisition functions
- New acquisition functions: `qSR`, `qNEI`, `LogEI`, `qLogEI`, `qLogNEI`
- `GammaPrior`, `HalfCauchyPrior`, `NormalPrior`, `HalfNormalPrior`, `LogNormalPrior`
and `SmoothedBoxPrior` classes
- Possibility to deserialize classes from optional class name abbreviations
- Basic deserialization tests using different class type specifiers
- Serialization user guide
- Environment variables user guide
- Utility for estimating memory requirements of discrete product search space
- `mypy` for search space and objectives

Changed
- Reorganized acquisition.py into `acquisition` subpackage
- Reorganized simulation.py into `simulation` subpackage
- Reorganized gaussian_process.py into `gaussian_process` subpackage
- Acquisition functions are now their own objects
- `acquisition_function_cls` constructor parameter renamed to `acquisition_function`
- User guide now explains the new objective classes
- Telemetry deactivation warning is only shown to developers
- `torch`, `gpytorch` and `botorch` are lazy-loaded for improved startup time
- If an exception is encountered during simulation, incomplete results are returned
with a warning instead of passing through the uncaught exception
- Environment variables `BAYBE_NUMPY_USE_SINGLE_PRECISION` and
`BAYBE_TORCH_USE_SINGLE_PRECISION` to enforce single point precision usage

Removed
- `model_params` attribute from `Surrogate` base class, `GaussianProcessSurrogate` and
`CustomONNXSurrogate`
- Dependency on `requests` package

Fixed
- `n_task_params` now evaluates to 1 if `task_idx == 0`
- Simulation no longer fails in `ignore` mode when lookup dataframe contains duplicate
parameter configurations
- Simulation no longer fails for targets in `MATCH` mode
- `closest_element` now works for array-like input of all kinds
- Structuring concrete subclasses no longer requires providing an explicit `type` field
- `_target(s)` attributes of `Objectives` are now (de-)serialized without leading
underscore to support user-friendly serialization strings
- Telemetry does not execute any code if it was disabled
- Running simulations no longer alters the states of the global random number generators

Deprecations
- The former `baybe.objective.Objective` class has been replaced with
`SingleTargetObjective` and `DesirabilityObjective`
- `acquisition_function_cls` constructor parameter for `BayesianRecommender`
- `VarUCB` and `qVarUCB` acquisition functions

Expired Deprecations (from 0.6.*)
- `BayBE` class
- `baybe.surrogate` module
- `baybe.targets.Objective` class
- `baybe.strategies.Strategy` class

Added
- Simulation user guide
- Example for transfer learning backtesting utility
- `pyupgrade` pre-commit hook
- Better human readable `__str__` representation of objective and targets
- Alternative dataframe deserialization from `pd.DataFrame` constructors

Changed
- More detailed and sophisticated search space user guide
- Support for Python 3.12
- Upgraded syntax to Python 3.9
- Bumped `onnx` version to fix vulnerability
- Increased threshold for low-dimensional GP priors
- Replaced `fit_gpytorch_mll_torch` with `fit_gpytorch_mll`
- Use `tox-uv` in pipelines

Fixed
- `telemetry` dependency is no longer a group (enables Poetry installation)