BAT.jl Files

BAT v4.0.0

Diff since v3.4.0

Breaking changes

Several algorithms have changed their names, but also their role:

• MCMCSampling has become TransformedMCMC.

• MetropolisHastings has become RandomWalk. It's parameters have changed (no deprecation for the parameter changes). Tuning and sample weighting scheme selection have moved to TransformedMCMC.

• PriorToGaussian has become PriorToNormal.

Partial deprecations are available for the above, so old code should run more or less unchanged (with deprecation warnings). Also:

• AdaptiveMHTuning has become AdaptiveAffineTuning, but is now used as a parameter for TransformedMCMC (formerly MCMCSampling) instead of RandomWalk (formerly MetropolisHastings).

• MCMCNoOpTuning has become NoMCMCTransformTuning.

• The arguments of HamiltonianMC have changed.

• MCMCTuningAlgorithm has been replaced by MCMCTransformTuning.

• The trafo parameter of algorithms has been renamed to pretransform, the trafo field in algorithm results has been renamed to f_pretransform.

• bat_report has been deprecated in favor of LazyReports.lazyreport (drop-in compatible).

New features

• Sampling, integration and mode-finding algorithms now generate a return value result = ..., evaluated::EvaluatedMeasure = ..., ...) if their target is a probability measure/distribution.

• The new RAMTuning is now the default (transform) tuning algorithm for RandomWalk (formerly MetropolisHastings). It typically results in a much faster burn-in process than AdaptiveAffineTuning (formerly AdaptiveMHTuning, the previous default).

• MCMC Sampling handles parameter scale and correlation adaptivity via via tunable space transformations instead of tuning covariance matrices in proposal distributions.

• MCMC tuning has been split into proposal tuning (algorithms of type MCMCProposalTuning) and transform turning (algorithms of type MCMCTransformTuning). Proposal tuning has now a much more limited role and often may be NoMCMCProposalTuning() (e.g. for RandomWalk).

• Added MGVISampling for Metric Gaussian Variational Inference.

Merged pull requests: - Add PolarShellDistribution (#475) (@oschulz) - CompatHelper: bump compat for AdvancedHMC in [weakdeps] to 0.8, (keep existing compat) (#480) (@github-actions[bot]) - Advanced HMC compat bump compatibility check (#481) (@Micki-D) - Improve dist transport implementation, support AffineDistribution (#482) (@oschulz)