Maintenance Release - Various Improvements across models, acquisition function optimization, PFN integration

Compatibility

• Require GPyTorch>=1.14.2 (#3055).

New Features

• Add EnsembleMapSaasSingleTaskGP (#3035, #3038, #3040).
• Allow different inferred noise levels for each task in MultitaskGP (#2997).
• Allow LatentKroneckerGP model to support different T values at train and test time (#3032, #3037).
• Allow qHypervolumeKnowledgeGradient to return log values for better numerical stability (#2974, #2976, #2979).
• Add NumericToCategoricalEncoding input transform (#2907).
• Added a MatheronPathModel - a DeterministicModel returning a Matheron path sample (#2984).
• Project points generated in acquisition function optization to the feasible space (#3010).
• Add support for non-uniform model weights in EnsembleModel and EnsemblePosterior (#2993).
• Allow optimizers to support negative indices for fixed features (#2970).
• Add worst known feasible value to constrained test problems (#3016).

Bug Fixes

• Fix optimize_acqf_mixed_alternating initialization with categorical features (#2986).
• Use IIDNormalSampler for PosteriorList by default to fix issue with correlated Sobol samples (#2977).
• Fix condition_on_observations to correctly apply input transforms and properly add data to train_inputs (#2989, #2990, #3034).
• Fix handling of input transforms for AdditiveMapSaasSingleTaskGP (#3042).
• Preserve train inputs and targets through transforms (#3044).
• Improve how qNEHVI handles pending points to avoid duplicate suggestions when initial pending points are passed (#2985).

Other Changes

• Add support for missing tasks in multi-task GP models (#2960).
• Add input constructor for LogConstrainedExpectedImprovement (#2973).
• Improve error handling and update documentation for inter-point constraints (#3003).
• Make AnalyticAcquisitionFunction._mean_and_sigma() return output dim consistently (#3028).
• Improve initialization with continuous relaxation in optimize_acqf_mixed_alternating (#3041).
• Implement ContextualDataset.__eq__() (#3005).
• Check shape of state dict when comparing input transforms (#3051).
• Add py.typed file to precent tools complainnig about type stubs (#2982).
• Improve best feasible objective computation; point user to use probability of feasibility (#3011).

Deprecations and removals

• Deprecate get_fitted_map_saas_ensemble() in favor of EnsembleMapSaasSingleTaskGP (#3036).

Changes to botorch_community

• Add MCAcquisition support to PFNModel (#3031).
• Add copula-based multivariate posterior for PFNModel (#3045).
• Allow PFNModel to load checkpoints from trainings done with automl/PFNs (#3017).
• Add support for Kaiming/He initialization for the VBLL mean (#3053).

v0.15.1: Compatibility release

[0.15.1] -- Aug 12, 2025

This is a compatibility release, coming only one week after 0.15.0.

New features

• Enable optimizing a sequence of acquisition functions in optimize_acqf (#2931).

Maintenance Release, Optimizer Improvements

New Features

• NP Regression Model w/ LIG Acquisition (#2683).
• Fully Bayesian Matern GP with dimension scaling prior (#2855).
• Option for input warping in non-linear fully Bayesian GPs (#2858).
• Support for condition_on_observations in FullyBayesianMultiTaskGP (#2871).
• Improvements to optimize_acqf_mixed_alternating:
  • Support categoricals in alternating optimization (#2866).
  • Batch mixed optimization (#2895).
  • Non-equidistant discrete dimensions for optimize_acqf_mixed_alternating (#2923).
  • Update syntax for categoricals in optimize_acqf_mixed_alternating (#2942).
  • Equality constraints for optimize_acqf_mixed_alternating (#2944).
• Multi-output acquisition functions and related utilities:
  • Multi-Output Acquisition Functions (#2935).
  • Utility for greedily selecting an approximate hypervolume maximizing subset (#2936).
  • Update optimize with NSGA-II (#2937).
  • Add utility for running pymoo NSGA-II (#2868).
• Batched L-BFGS-B for more efficient acquisition function optimization (#2870, #2892).
• Pathwise Thompson sampling for ensemble models (#2877).
• ROBOT tutorial notebook (#2883).
• Add community notebooks to the botorch.org website (#2913).

Bug Fixes

• Fix model paths in prior fitted networks (#2843).
• Fix a bug where input transforms were not applied in fully Bayesian models in train mode (#2859).
• Fix local Y vs global Y_Train in generate_batch function in TURBO tutorial (#2862).
• Fix CUDA support for FullyBayesianMTGP (#2875).
• Fix edge case with NaNs in is_non_dominated (#2925).
• Normalize for correct fidelity in qLowerBoundMaxValueEntropy (#2930).
• Bug: Botorch_community VBLLModel posterior doesn't work with single value tensor (#2929).
• Fix variance shape bug in Riemann posterior (#2939).
• Fix input constructor for LogProbabilityOfFeasibility (#2945).
• Fix AugmentedRosenbrock problem and expand testing for optimizers (#2950).

Other Changes

• Improved documentation for optimize_acqf (#2865).
• Fully Bayesian Multi-Task GP cleanup (#2869).
• average_over_ensemble_models decorator for acquisition functions (#2873).
• Changes to I-BNN tutorial (#2889).
• Allow batched fixed features in gen_candidates_scipy and gen_candidates_torch (#2893)
• Refactor of MultiTask / FullyBayesianMultiTaskGP to use ProductKernel and IndexKernel (#2908).
• Various changes to PFNs to improve Ax compatibility (#2915, #2940).
• Eliminate expensive indexing in separate_mtmvn (#2920).
• Added reset method to StoppingCriterion (#2927).
• Simplify closure dispatch (#2947).
• Add BaseTestProblem.is_minimization_problem property (#2949).
• Simplify NdarrayOptimizationClosure (#2951).

Maintenance Release, PFN integration, VBLL surrogates, Classifier-based constraint support

Highlights

• Prior Fitted Network (PFN) surrogate model integration (#2784).
• Variational Bayesian last-layer models as surrogate Models (#2754).
• Probabilities of feasibility for classifier-based constraints in acquisition functions (#2776).

New Features

• Helper for evaluating feasibility of candidate points (#2733).
  • Check for feasibility in gen_candidates_scipy and error out for infeasible candidates (#2737).
  • Return a feasible candidate if there is one and return_best_only=True (#2778).
• Allow for observation noise without provided evaluation_mask mask in ModelListGP (#2735).
• Implement incremental qLogNEI via incremental argument to qLogNoisyExpectedImprovement (#2760).
• Add utility for computing AIC/BIC/MLL from a model (#2785).
• New test functions:
  • Multi-fidelity test functions with discrete fidelities (#2796).
  • Keane bump function (#2802).
  • Mixed Ackley test function (#2830).
  • LABS test function (#2832).
• Add parameter types to test functions to support problems defined in mixed / discrete spaces (#2809).
  • Add input validation to test functions (#2829).
• Add [q]LogProbabilityOfFeasibility acquisition functions (#2815).

Bug Fixes

• Remove hard-coded dtype from best_f buffers (#2725).
• Fix dtype/nan issue in StratifiedStandardize (#2757).
• Properly handle observed noise in AdditiveMapSaasSingleTaskGP with outcome transforms (#2763).
• Do not count STOPPED (due to specified budget) as a model fitting failure (#2767).
• Ensure that initialize_q_batch always includes the maximum value when called in batch mode (#2773).
• Fix posterior with observation noise in batched MTGP models (#2782).
• Detach tensor in gen_candidates_scipy to avoid test failure due to new warning (#2797).
• Fix batch computation in Pivoted Cholesky (#2823).

Other Changes

• Add optimal values for synthetic contrained optimization problems (#2730).
  • Update max_hv and reference point for Penicillin problem (#2771).
  • Add optimal value to SpeedReducer problem (#2799).
• Update nonlinear_constraint_is_feasible to return a boolean tensor (#2731).
• Restructure sampling methods for info-theoretic acquisition functions (#2753).
• Prune baseline points in qLogNEI by default (#2762).
• Misc updates to MES-based acqusition functions (#2769).
• Pass option to reset submodules in train method for fully Bayesian models (#2783).
• Put outcome transforms into train mode in model constructors (#2817).
• LogEI: select cache_root based on model support (#2820).
• Remove Ax dependency from BoTorch tutorials and reference Ax tutorials instead (#2839).

Deprecations and removals

• Remove deprecated gp_sampling module (#2768).
• Remove qMultiObjectiveMaxValueEntropy acquisition function (#2800).
• Remove model converters (#2801).

Maintenance Release, Website Upgrade, BO with Relevance Pursuit, LatentKroneckerGP and MAP-SAAS Models

Highlights

• BoTorch website has been upgraded to utilize Docusaurus v3, with the API
  reference being hosted by ReadTheDocs. The tutorials now expose an option to
  open with Colab, for easy access to a runtime with modifiable tutorials.
  The old versions of the website can be found at archive.botorch.org (#2653).
• RobustRelevancePursuitSingleTaskGP, a robust Gaussian process model that adaptively identifies
  outliers and leverages Bayesian model selection (paper) (#2608, #2690, #2707).
• LatentKroneckerGP, a scalable model for data on partially observed grids, like the joint modeling
  of hyper-parameters and partially completed learning curves in AutoML (paper) (#2647).
• Add MAP-SAAS model, which utilizes the sparse axis-aligned subspace priors
  (paper) with MAP model fitting (#2694).

Compatibility

• Require GPyTorch==1.14 and linear_operator==0.6 (#2710).
• Remove support for anaconda (official package) (#2617).
• Remove mpmath dependency pin (#2640).
• Updates to optimization routines to support SciPy>1.15:
  • Use threadpoolctl in minimize_with_timeout to prevent CPU oversubscription (#2712).
  • Update optimizer output parsing to make model fitting compatible with SciPy>1.15 (#2667).

New Features

• Add support for priors in OAK Kernel (#2535).
• Add BatchBroadcastedTransformList, which broadcasts a list of InputTransforms over batch shapes (#2558).
• InteractionFeatures input transform (#2560).
• Implement percentile_of_score, which takes inputs data and score, and returns the percentile of
  values in data that are below score (#2568).
• Add optimize_acqf_mixed_alternating, which supports optimization over mixed discrete & continuous spaces (#2573).
• Add support for PosteriorTransform to get_optimal_samples and optimize_posterior_samples (#2576).
• Support inequality constraints & X_avoid in optimize_acqf_discrete (#2593).
• Add ability to mix batch initial conditions and internal IC generation (#2610).
• Add qPosteriorStandardDeviation acquisition function (#2634).
• TopK downselection for initial batch generation. (#2636).
• Support optimization over mixed spaces in optimize_acqf_homotopy (#2639).
• Add InfeasibilityError exception class (#2652).
• Support InputTransforms in SparseOutlierLikelihood and get_posterior_over_support (#2659).
• StratifiedStandardize outcome transform (#2671).
• Add center argument to Normalize (#2680).
• Add input normalization step in Warp input transform (#2692).
• Support mixing fully Bayesian & SingleTaskGP models in ModelListGP (#2693).
• Add abstract fully Bayesian GP class and fully Bayesian linear GP model (#2696, #2697).
• Tutorial on BO constrained by probability of classification model (#2700).

Bug Fixes

• Fix error in decoupled_mobo tutorial due to torch/numpy issues (#2550).
• Raise error for MTGP in batch_cross_validation (#2554).
• Fix posterior method in BatchedMultiOutputGPyTorchModel for tracing JIT (#2592).
• Replace hard-coded double precision in test_functions with default dtype (#2597).
• Remove as_tensor argument of set_tensors_from_ndarray_1d (#2615).
• Skip fixed feature enumerations in optimize_acqf_mixed that can't satisfy the parameter constraints (#2614).
• Fix get_default_partitioning_alpha for >7 objectives (#2646).
• Fix random seed handling in sample_hypersphere (#2688).
• Fix bug in optimize_objective with fixed features (#2691).
• FullyBayesianSingleTaskGP.train should not return None (#2702).

Other Changes

• More efficient sampling from KroneckerMultiTaskGP (#2460).
• Update HigherOrderGP to use new priors & standardize outcome transform by default (#2555).
• Update initialize_q_batch methods to return both candidates and the corresponding acquisition values (#2571).
• Update optimization documentation with LogEI insights (#2587).
• Make all arguments in optimize_acqf_homotopy explicit (#2588).
• Introduce trial_indices argument to SupervisedDataset (#2595).
• Make optimizers raise an error when provided negative indices for fixed features (#2603).
• Make input transforms Modules by default (#2607).
• Reduce memory usage in ConstrainedMaxPosteriorSampling (#2622).
• Add clone method to datasets (#2625).
• Add support for continuous relaxation within optimize_acqf_mixed_alternating (#2635).
• Update indexing in qLogNEI._get_samples_and_objectives to support multiple input batches (#2649).
• Pass X to OutcomeTransforms (#2663).
• Use mini-batches when evaluating candidates within optimize_acqf_discrete_local_search (#2682).

Deprecations

• Remove HeteroskedasticSingleTaskGP (#2616).
• Remove FixedNoiseDataset (#2626).
• Remove support for legacy format non-linear constraints (#2627).
• Remove maximize option from information theoretic acquisition functions (#2590).

Increased robustness to dimensionality with updated hyperparameter priors

[0.12.0] -- Sep 17, 2024

Major changes

• Update most models to use dimension-scaled log-normal hyperparameter priors by
  default, which makes performance much more robust to dimensionality. See
  discussion #2451 for details. The only models that are not changed are those
  for fully Bayesian models and PairwiseGP; for models that utilize a
  composite kernel, such as multi-fidelity/task/context, this change only
  affects the base kernel (#2449, #2450, #2507).
• Use Standarize by default in all the models using the upgraded priors. In
  addition to reducing the amount of boilerplate needed to initialize a model,
  this change was motivated by the change to default priors, because the new
  priors will work less well when data is not standardized. Users who do not
  want to use transforms should explicitly pass in None (#2458, #2532).

Compatibility

• Unpin NumPy (#2459).
• Require PyTorch>=2.0.1, GPyTorch==1.13, and linear_operator==0.5.3 (#2511).

New features

• Introduce PathwiseThompsonSampling acquisition function (#2443).
• Enable qBayesianActiveLearningByDisagreement to accept a posterior
  transform, and improve its implementation (#2457).
• Enable SaasPyroModel to sample via NUTS when training data is empty (#2465).
• Add multi-objective qBayesianActiveLearningByDisagreement (#2475).
• Add input constructor for qNegIntegratedPosteriorVariance (#2477).
• Introduce qLowerConfidenceBound (#2517).
• Add input constructor for qMultiFidelityHypervolumeKnowledgeGradient (#2524).
• Add posterior_transform to ApproximateGPyTorchModel.posterior (#2531).

Bug fixes

• Fix batch_shape default in OrthogonalAdditiveKernel (#2473).
• Ensure all tensors are on CPU in HitAndRunPolytopeSampler (#2502).
• Fix duplicate logging in generation/gen.py (#2504).
• Raise exception if X_pending is set on the underlying AcquisitionFunction
  in prior-guided AcquisitionFunction (#2505).
• Make affine input transforms error with data of incorrect dimension, even in
  eval mode (#2510).
• Use fidelity-aware current_value in input constructor for qMultiFidelityKnowledgeGradient (#2519).
• Apply input transforms when computing MLL in model closures (#2527).
• Detach fval in torch_minimize to remove an opportunity for memory leaks
  (#2529).

Documentation

• Clarify incompatibility of inter-point constraints with get_polytope_samples
  (#2469).
• Update tutorials to use the log variants of EI-family acquisition functions,
  don't make tutorials pass Standardize unnecessarily, and other
  simplifications and cleanup (#2462, #2463, #2490, #2495, #2496, #2498, #2499).
• Remove deprecated FixedNoiseGP (#2536).

Other changes

• More informative warnings about failure to standardize or normalize data
  (#2489).
• Suppress irrelevant warnings in qHypervolumeKnowledgeGradient helpers
  (#2486).
• Cleaner botorch/acquisition/multi_objective directory structure (#2485).
• With AffineInputTransform, always require data to have at least two
  dimensions (#2518).
• Remove deprecated argument data_fidelity to SingleTaskMultiFidelityGP and
  deprecated model FixedNoiseMultiFidelityGP (#2532).
• Raise an OptimizationGradientError when optimization produces NaN gradients (#2537).
• Improve numerics by replacing torch.log(1 + x) with torch.log1p(x)
  and torch.exp(x) - 1 with torch.special.expm1 (#2539, #2540, #2541).

Maintenance Release, I-BNN Kernel

Compatibility

• Pin NumPy to <2.0 (#2382).
• Require GPyTorch 1.12 and LinearOperator 0.5.2 (#2408, #2441).

New features

• Support evaluating posterior predictive in MultiTaskGP (#2375).
• Infinite width BNN kernel (#2366) and the corresponding tutorial (#2381).
• An improved elliptical slice sampling implementation (#2426).
• Add a helper for producing a DeterministicModel using a Matheron path (#2435).

Deprecations and Deletions

• Stop allowing some arguments to be ignored in acqf input constructors (#2356).
• Reap deprecated **kwargs argument from optimize_acqf variants (#2390).
• Delete DeterministicPosterior and DeterministicSampler (#2391, #2409, #2410).
• Removed deprecated CachedCholeskyMCAcquisitionFunction (#2399).
• Deprecate model conversion code (#2431).
• Deprecate gp_sampling module in favor of pathwise sampling (#2432).

Bug Fixes

• Fix observation noise shape for batched models (#2377).
• Fix sample_all_priors to not sample one value for all lengthscales (#2404).
• Make (Log)NoisyExpectedImprovement create a correct fantasy model with
  non-default SingleTaskGP (#2414).

Other Changes

• Various documentation improvements (#2395, #2425, #2436, #2437, #2438).
• Clean up **kwargs arguments in qLogNEI (#2406).
• Add a NumericsWarning for Legacy EI implementations (#2429).

Maintenance Release

New Features

• Implement qLogNParEGO (#2364).
• Support picking best of multiple fit attempts in fit_gpytorch_mll (#2373).

Deprecations

• Many functions that used to silently ignore arbitrary keyword arguments will now
  raise an exception when passed unsupported arguments (#2327, #2336).
• Remove UnstandardizeMCMultiOutputObjective and UnstandardizePosteriorTransform (#2362).

Bug Fixes

• Remove correlation between the step size and the step direction in sample_polytope (#2290).
• Fix pathwise sampler bug (#2337).
• Explicitly check timeout against None so that 0.0 isn't ignored (#2348).
• Fix boundary handling in sample_polytope (#2353).
• Avoid division by zero in normalize & unnormalize when lower & upper bounds are equal (#2363).
• Update sample_all_priors to support wider set of priors (#2371).

Other Changes

• Clarify is_non_dominated behavior with NaN (#2332).
• Add input constructor for qEUBO (#2335).
• Add LogEI as a baseline in the TuRBO tutorial (#2355).
• Update polytope sampling code and add thinning capability (#2358).
• Add initial objective values to initial state for sample efficiency (#2365).
• Clarify behavior on standard deviations with <1 degree of freedom (#2357).

Maintenance Release, SCoreBO

Compatibility

• Reqire Python >= 3.10 (#2293).

New Features

• SCoreBO and Bayesian Active Learning acquisition functions (#2163).

Bug Fixes

• Fix non-None constraint noise levels in some constrained test problems (#2241).
• Fix inverse cost-weighted utility behaviour for non-positive acquisition values (#2297).

Other Changes

• Don't allow unused keyword arguments in Model.construct_inputs (#2186).
• Re-map task values in MTGP if they are not contiguous integers starting from zero (#2230).
• Unify ModelList and ModelListGP subset_output behavior (#2231).
• Ensure mean and interior_point of LinearEllipticalSliceSampler have correct shapes (#2245).
• Speed up task covariance of LCEMGP (#2260).
• Improvements to batch_cross_validation, support for model init kwargs (#2269).
• Support custom all_tasks for MTGPs (#2271).
• Error out if scipy optimizer does not support bounds / constraints (#2282).
• Support diagonal covariance root with fixed indices for LinearEllipticalSliceSampler (#2283).
• Make qNIPV a subclass of AcquisitionFunction rather than AnalyticAcquisitionFunction (#2286).
• Increase code-sharing of LCEMGP & define construct_inputs (#2291).

Deprecations

• Remove deprecated args from base MCSampler (#2228).
• Remove deprecated botorch/generation/gen/minimize (#2229).
• Remove fit_gpytorch_model (#2250).
• Remove requires_grad_ctx (#2252).
• Remove base_samples argument of GPyTorchPosterior.rsample (#2254).
• Remove deprecated mvn argument to GPyTorchPosterior (#2255).
• Remove deprecated Posterior.event_shape (#2320).
• Remove **kwargs & deprecated indices argument of Round transform (#2321).
• Remove Standardize.load_state_dict (#2322).
• Remove FixedNoiseMultiTaskGP (#2323).