Botorch

* Require linear_operator == 0.5.1 (1963).

Compatibility
* Require Python >= 3.9.0 (1924).
* Require PyTorch >= 1.13.1 (1960).
* Require linear_operator == 0.5.0 (1961).
* Require GPyTorch == 1.11 (1961).

Highlights
* Introduce `OrthogonalAdditiveKernel` (1869).
* Speed up LCE-A kernel by over an order of magnitude (1910).
* Introduce `optimize_acqf_homotopy`, for optimizing acquisition functions with homotopy (1915).
* Introduce `PriorGuidedAcquisitionFunction` (PiBO) (1920).
* Introduce `qLogExpectedImprovement`, which provides more accurate numerics than `qExpectedImprovement` and can lead to significant optimization improvements (1936).
* Similarly, introduce `qLogNoisyExpectedImprovement`, which is analogous to `qNoisyExpectedImprovement` (1937).

New Features
* Add constrained synthetic test functions `PressureVesselDesign`, `WeldedBeam`, `SpeedReducer`, and `TensionCompressionString` (1832).
* Support decoupled fantasization (1853) and decoupled evaluations in cost-aware utilities (1949).
* Add `PairwiseBayesianActiveLearningByDisagreement`, an active learning acquisition function for PBO and BOPE (1855).
* Support custom mean and likelihood in `MultiTaskGP` (1909).
* Enable candidate generation (via `optimize_acqf`) with both `non_linear_constraints` and `fixed_features` (1912).
* Introduce `L0PenaltyApproxObjective` to support L0 regularization (1916).
* Enable batching in `PriorGuidedAcquisitionFunction` (1925).

Other changes
* Deprecate `FixedNoiseMultiTaskGP`; allow `train_Yvar` optionally in `MultiTaskGP` (1818).
* Implement `load_state_dict` for SAAS multi-task GP (1825).
* Improvements to `LinearEllipticalSliceSampler` (1859, 1878, 1879, 1883).
* Allow passing in task features as part of X in MTGP.posterior (1868).
* Improve numerical stability of log densities in pairwise GPs (1919).
* Python 3.11 compliance (1927).
* Enable using constraints with `SampleReducingMCAcquisitionFunction`s when using `input_constructor`s and `get_acquisition_function` (1932).
* Enable use of `qLogExpectedImprovement` and `qLogNoisyExpectedImprovement` with Ax (1941).

Bug Fixes
* Enable pathwise sampling modules to be converted to GPU (1821).
* Allow `Standardize` modules to be loaded once trained (1874).
* Fix memory leak in Inducing Point Allocators (1890).
* Correct einsum computation in `LCEAKernel` (1918).
* Properly whiten bounds in MVNXPB (1933).
* Make `FixedFeatureAcquisitionFunction` convert floats to double-precision tensors rather than single-precision (1944).
* Fix memory leak in `FullyBayesianPosterior` (1951).
* Make `AnalyticExpectedUtilityOfBestOption` input constructor work correctionly with multi-task GPs (1955).

New Features
* Support inferred noise in `SaasFullyBayesianMultiTaskGP` (1809).

Other Changes
* More informative error message when `Standardize` has wrong batch shape (1807).
* Make GIBBON robust to numerical instability (1814).
* Add `sample_multiplier` in EUBO's `acqf_input_constructor` (1816).

Bug Fixes
* Only do checks for `_optimize_acqf_sequential_q` when it will be used (1808).
* Fix an issue where `PairwiseGP` comparisons might be implicitly modified (1811).

Compatibility
* Require GPyTorch == 1.10 and linear_operator == 0.4.0 (1803).

New Features
* Polytope sampling for linear constraints along the q-dimension (1757).
* Single-objective joint entropy search with additional conditioning, various improvements to entropy-based acquisition functions (1738).

Other changes
* Various updates to improve numerical stability of `PairwiseGP` (1754, 1755).
* Change batch range for `FullyBayesianPosterior` (1176a38352b69d01def0a466233e6633c17d6862, 1773).
* Make `gen_batch_initial_conditions` more flexible (1779).
* Deprecate `objective` in favor of `posterior_transform` for `MultiObjectiveAnalyticAcquisitionFunction` (1781).
* Use `prune_baseline=True` as default for `qNoisyExpectedImprovement` (1796).
* Add `batch_shape` property to `SingleTaskVariationalGP` (1799).
* Change minimum inferred noise level for `SaasFullyBayesianSingleTaskGP` (1800).

Bug fixes
* Add `output_task` to `MultiTaskGP.construct_inputs` (1753).
* Fix custom bounds handling in test problems (1760).
* Remove incorrect `BotorchTensorDimensionWarning` (1790).
* Fix handling of non-Container-typed positional arguments in `SupervisedDatasetMeta` (1663).

New Features
* Add BAxUS tutorial (1559).

Other changes
* Various improvements to tutorials (1703, 1706, 1707, 1708, 1710, 1711, 1718, 1719, 1739, 1740, 1742).
* Allow tensor input for `integer_indices` in `Round` transform (1709).
* Expose `cache_root` in qNEHVI input constructor (1730).
* Add `get_init_args` helper to `Normalize` & `Round` transforms (1731).
* Allowing custom dimensionality and improved gradient stability in `ModifiedFixedSingleSampleModel` (1732).

Bug fixes
* Improve batched model handling in `_verify_output_shape` (1715).
* Fix qNEI with Derivative Enabled BO (1716).
* Fix `get_infeasible_cost` for objectives that require X (1721).

Compatibility
* Require PyTorch >= 1.12 (1699).

New Features
* Introduce pathwise sampling API for efficiently sampling functions from (approximate) GP priors and posteriors (1463).
* Add `OneHotToNumeric` input transform (1517).
* Add `get_rounding_input_transform` utility for constructing rounding input transforms (1531).
* Introduce `EnsemblePosterior` (1636).
* Inducing Point Allocators for Sparse GPs (1652).
* Pass `gen_candidates` callable in `optimize_acqf` (1655).
* Adding `logmeanexp` and `logdiffexp` numerical utilities (1657).

Other changes
* Warn if inoperable keyword arguments are passed to optimizers (1421).
* Add `BotorchTestCase.assertAllClose` (1618).
* Add `sample_shape` property to `ListSampler` (1624).
* Do not filter out `BoTorchWarning`s by default (1630).
* Introduce a `DeterministicSampler` (1641).
* Warn when optimizer kwargs are being ignored in BoTorch optim utils `_filter_kwargs` (1645).
* Don't use `functools.lru_cache` on methods (1650).
* More informative error when someone adds a module without updating the corresponding rst file (1653).
* Make indices a buffer in `AffineInputTransform` (1656).
* Clean up `optimize_acqf` and `_make_linear_constraints` (1660, 1676).
* Support NaN `max_reference_point` in `infer_reference_point` (1671).
* Use `_fast_solves` in `HOGP.posterior` (1682).
* Approximate qPI using `MVNXPB` (1684).
* Improve filtering for `cache_root` in `CachedCholeskyMCAcquisitionFunction` (1688).
* Add option to disable retrying on optimization warning (1696).

Bug fixes
* Fix normalization in Chebyshev scalarization (1616).
* Fix `TransformedPosterior` missing batch shape error in `_update_base_samples` (1625).
* Detach `coefficient` and `offset` in `AffineTransform` in eval mode (1642).
* Fix pickle error in `TorchPosterior` (1644).
* Fix shape error in `optimize_acqf_cyclic` (1648).
* Fixed bug where `optimize_acqf` didn't work with different batch sizes (1668).
* Fix EUBO optimization error when two Xs are identical (1670).
* Bug fix: `_filter_kwargs` was erroring when provided a function without a `__name__` attribute (1678).