Skip to content

Latest commit

 

History

History
105 lines (83 loc) · 5.03 KB

NEWS.md

File metadata and controls

105 lines (83 loc) · 5.03 KB
Raw

News

Version 0.12.0 (2020-04-09)

  • Optimisation interface now matches that of the Optim.jl package.
  • Added support for heteroscedastic noise in the exact inference case.
  • Added autodifferentiation support for kernels with fixed parameters.
  • Bug fixes in elliptical slice sampler.

Version 0.11.0 (2019-11-08)

  • Introduced an alternative sampling method to Hamiltonian Monte-Carlo, namely an elliptical slice sampler
    • Performing inference via an ESS is more robust to poor hyperparameter initialisation
    • An ESS is often able to explore the posterior space more efficiently in the case of highly dependent Gaussian variables.

Version 0.10.0 (2019-09-21)

  • Introduced sparse approximation methods
    • Subset of regressors, deterministic training conditional, fully independent training conditional and full-scale approximation are all available
  • Extended functionality to include leave-one-out cross-validation
  • Introduced functionality to enable variational inference in GPs with non-Gaussian data.
    • The approach used is a variant of that presented in Khan et. al.
    • Currently limited to only Poisson data. Additional likelihood functionality will be added soon.
  • Deprecated GPMC in favor of GPA. This is to be inline with that fact that approximate inference in a GP is not limited to MCMC, but variational methods can now be used.

Version 0.9.0 (2018-12-07)

  • Introduction of ElasticGPE to allow GP which can grow without "refitting" whole Gaussian process (see also #88).
  • Various performance and interface improvements to Kernels
  • Introduction of ADKernel to simplify introduction of new/custom kernels through use of auto-differentiation.

Version 0.8.0 (2018-10-02)

  • Updated Julia requirement to v1.0
  • Updated requirement of RecipesBase.jl to v0.6

Note: Could not create release compatible with both Julia v0.7 and Julia v1.0 due to RecipesBase.jl dependency

Version 0.7.0 (2018-09-28)

  • Updated Julia requirement to v0.7
  • Performance improvements to Kernels
  • Added dependencies on StatsFuns and SpecialFunctions
  • Removed dependency on Compat
  • Renamed FixedKern to FixedKernel
  • Added type parameters to GPE, GPMC, ProdKernel, and SumKernel, ProdMean and SumMean
  • Renamed fields of GPE and GPE (x instead of X, mean instead of m, kernel instead of k, and nobs instead of nobsv)
  • Renamed fields of FixedKernel and Masked (kernel instead of kern)
  • Renamed fields of ProdKernel and SumKernel (kernels instead of kerns)
  • Renamed keyword arguments of the GPE constructor to kernel and mean
  • Renamed function subkernels and submeans to components

Version 0.6.0 (2018-03-02)

  • Updated optimization code to be compatible with new Optim.jl API
  • Removed Klara dependency
  • Performance improvements to predict functions

Version 0.5.0 (2017-10-18)

  • Updated Julia version requirement to 0.6
  • GP type has been renamed to GPE (GP exact) for Gaussian likelihoods
  • Introduced GPMC type for fitting models with non-Gaussian likelihoods:
    • Bernouilli, Poisson, Binomial and student-t likelihoods available
  • Introduced priors for parameters of the kernal, mean, and likelihood functions
  • MCMC available for GPE and GPMC type
  • Changed plotting functions to use Plots.jl
  • Created notebooks illustrating package features

Version 0.4.0 (2016-10-04)

  • Julia requirement moved up to version 0.5
  • Major speed improvements for fitting of GP object, and for covariance and gradient calculations
  • New Masked kernel
  • Various bug fixes

Version 0.3.0 (2016-07-11)

  • Introduced KernelData type to recycle calculations
  • Removed Winston plotting functions and implemented PyPlot as an alternative
  • Created methods for mean and cov functions of the Mean and Kernel objects
  • Fixed optimize! function to be consistent with most recent version of Optim.jl
  • Improvements to the Periodic kernel
  • fit! function no longer exported due to clash with a few packages

Version 0.2.1 (2016-06-06)

  • Added fit! function to fit a new set observations to existing GP object

Version 0.2.0 (2016-06-03)

  • Julia requirement moved up to v0.4
  • Support added for ScikitLearn
  • rand and rand! functions added to sample prior and posterior paths of Gaussian process
  • Major speed improvements for gradient calculations of stationary ARD kernels
  • Minor fixes for some kernels

Version 0.1.4 (2015-10-28)

  • Fixed plotting deprecation errors with Julia 0.4

Version 0.1.3 (2015-10-26)

  • Major speed improvements to kernel calculations, in particular to stationary and composite kernels
  • Fixed depraction warnings for Julia v0.4
  • All stationary kernels have the super type Stationary
  • Distance matrix calculations outsourced to Distances

Version 0.1.2 (2015-06-04)

  • Improvements in speed for predict and fitting functions
  • Positive definite matrix calculations outsourced to PDMats