Release 3.0.0

Breaking changes
This release includes a minor breaking change:

The Record, FrozenRecord and OptimizationResult classes are now also generic in the model type. Any type-annotated code using these may need to be updated to include the model type annotations as appropriate. Runtime behaviour is unaffected.

New features
Customisable initial point selection for optimization (#808)
Customisable handling of model updates in AskTellOptimizer (including an AskTellOptimizerNoTraining class for use with non-trainable models) (#815)
DeepEnsemble support for pass-through keras compile arguments (#816)

Improvements/fixes
Support for tensorflow 2.15 (#819)
Filter out local datasets when calling base rule (#805)
Speed up get_unique_points_mask (used by BatchTrustRegionBox) (#813)

Documentation
README and CONTRIBUTING updates (#804, #802, #810)
Mixed search space notebook (#818)

New Contributors: @nfergu (#816)

Full Changelog: v2.0.0...v3.0.0

Release 2.0.0

Breaking changes
This release includes a few minor breaking changes:

1. ProbabilisticModel is now a pure interface (#797)

• ProbabilisticModel.log is now an abstract method and needs to be explicitly specified in any concrete model class implementation (though it can be empty)
• ProbabilisticModel.get_module_with_variables is now a utility function in trieste.models.utils

2. TREGO (TrustRegion) and TURBO were reimplemented using the new batch-trust-region classes (#778, #791)

• a TURBO rule must now be initialised as BatchTrustRegionBox(TURBOBox(search_space)) instead of TURBO(search_space)
• a TREGO rule must now be initialised as BatchTrustRegionBox(TREGOBox(search_space)) instead of TrustRegion()
• the internal State structures exposed by these rules are now also different: BatchTrustRegion.State instead of Turbo.State or TrustRegion.State, with additional values such as eps accessible via the subspaces in rule._subspaces instead.

New features
Multi trust region acquisition rules (#773, #777, #778, #783)
Local models and datasets (#788, #791)
Expose model optimization result in optimize method and get_last_optimization_result function (#774, #797)

Improvements/fixes
Stop trajectory sampling ignoring active-dims in the kernel (#790)
Stop randomize_hyperparameters generating repeating values (#785)
Handle unconstrained priors in randomize_hyperparameters (#796)
Support optimization of multiple points in batch spaces (#787)
Allow Boxes with zero width (#780)
Deepcopy search spaces (#776)
Use int64 when calculating data size in split_acquisition_function (#795)
Start using check_shape for shape checking (#770)
Cleanup tutorials (#769, #771)

Build changes
Parallelise integration test run (#775)
Integration test fragility (#786, #798)

Full Changelog: v1.2.0...v2.0.0

Release 1.2.0

New features
SeparateIndependent kernel support (#711, #719, #736, #745, #747, #753)
TuRBO acquisition rule (#739, #762)
Saving models as tf.saved_model (#750)

Improvements
Fix deepcopy of posterior cache and VGP parameters (#741, #752)
Fix learning rate reset error with gpflux and keras models (#740)
Support Tensorflow 2.11 and 2.12 (#746)
Make IndependentReparametrizationSampler support XLA (#718, #748)
Handle broadcasting in DeepEnsemble models (#727)
Support calling qmc_normal_samples with tensors (#723)
Make default tensorboard metric names unique (#726)
Make Reducer AFs generic in model type and add Map reducer (#759)
Add continue_optimization method to BayesianOptimizer (#755)
Additional test problems (#744)

Build changes
Add default PR template (#756)
Support installation from pypi on MacOS (#730)
Get notebooks to build with Python 3.10 (#728, #732, #734, #738, #742)
Parallelise test runs (#743, #749)
Reduce notebook file sizes (#721)

Full Changelog: v1.1.2...v1.2.0

Release 1.1.2

Bug fixes
Support disabling skip in QMC sampling (#715)

Documentation
REMBO notebook (#710)

Full Changelog: v1.1.1...v1.1.2

Release 1.1.1

Bug fixes
Fix Batch QMC sampling (#713)

Full Changelog: v1.1.0...v1.1.1

Release 1.1.0

New features
MUMBO: MUlti-task Max-value Bayesian Optimization (#699)
Support QMC sampling in reparameterization samplers (#708)

Improvements
Fix SVGP update equations (#709)
Support tf compilation in MultifidelityAutoregressive and DeepEnsemble (#691, #698)

Build changes
New README.md (#690, #705, #706)
Track code coverage (#702)
Unpin mypy and black versions (#701)
Test against Python 3.10 (#688, #696)
Work around docs issue with setuptools+gym (#695)

New contributors
@eltociear made their first contribution in #700

Full Changelog: v1.0.0...v1.1.0

Release 1.0.0

Note: this release marks the first 1.x release but is compatible with 0.13.3. Future releases will (try to) conform to semantic versioning. Since trieste is a research-led toolbox, this may result in reasonably frequent major version increments.

New features
NARGP multifidelity model (#665)
BO-specific inducing point allocators (#683)

Improvements
Support for explicit constraints in ExpectedImprovement (#664)
Support broadcasting in search space contains method (#677)
Improved logging for GPflow models (#680)
Faster sampler for deep ensembles (#682)

Build changes
.gitignore additions (#679)
Upgrade to GPflow 2.7.0 (#684)
Workaround for slowtest OOM crash (#685)

Full Changelog: v0.13.3...v1.0.0

Release 0.13.3

This release reintroduces ProbabilityOfImprovement (#638) and Pareto diverse sample method (#643), which were temporarily removed in 0.13.2.

Note that the minimum supported TensorFlow version has been raised to 2.5.

New features
Batch Expected Improvement (#641, #653)
Portfolio method for Batch BO (#651, #659, #663)
Multifidelity modeling (#621, #654)
Explicit constraints (#656, #660)

Improvements
Allow scipy optimizer to be changed (#655)
Allow arbitrary dataset/model tags (#668) — note that this may break type checking for existing code
Slight improvements to synthetic objective functgions (#671)

Build changes
Support latest gpflow and gpflux (#649)
Support more recent tox versions (#669)

Full Changelog: v0.13.1...v0.13.3

Release 0.13.2

This release fixes the 0.13.1 release by temporarily removing the two new features (#638 and #643) as they were preventing trieste from being used with the latest release of GPflow. A future release will reintroduce them.

The fix to handle constant priors in randomize_hyperparameters (#646) remains in the release.

Full Changelog: v0.13.0...v0.13.2

Release 0.13.1

New features
ProbabilityOfImprovement acquisition function (#638)
Pareto diverse sample method (#643)

Improvements
Handle constant priors in randomize_hyperparameters (#646)

Full Changelog: v0.13.0...v0.13.1