v0.7.2

Bugfix release to fix a few outstanding issues. Fixes issues #241 and #242. Changes in this release:

• Update to Dimension Reduction to improve memory usage with some additional tests
• Update to documentation to reflect new dependency requirements
• Improvements to GPU build script in setup.py to remove outdated shared object library files.

v0.7.1

Bug fix release to correct the GP demos to ensure that they run with the new
implementation. Changes include:

• Update Patsy integration to allow using formulae with LHS
• Updates to demos gp_demos.py, gp_demo.R, excalibur_workshop_demo.py, gp_kernel_demos.py, multioutput_tutorial.py
• Add new printing function to the projectile code to replace previous, modified above demos to use this.

v0.7.0

New release incorporating recent work. New features in this release:

• Maxi-min LHC Experimental Designs
• Mulit-output history matching
• Emulator validation
• GPU implementation now supports Priors
• Additional fixes and improvements to documentation and examples

v0.6.1

Bugfix release addressing a few issues identified when using the new version of the GP. New in this release:

• GP Predictions using mean functions have been corrected to fix an error where the mean function was not correctly subtracted from the targets when fitting.
• Fix a few problems caused by patsy objects not supporting pickling with parallel fitting/predictions with multiplt outputs.

v0.6.0

New release with major improvements in CPU and GPU versions of the GP class.

Changes in CPU version:

• Mean functions are implemented using patsy to form design matrices.
• Default Prior distributions chosen for correlation lengths and nugget (if fitted).
• When fitting, starting point of optimization is a draw from the prior.
• New parameters class and coordinate transformation objects to make parameter meanings more transparent.
• Mean priors are now multivariate normal distributions.
• Mean functions are fit analytically.
• Additional kernels include Squared Exponential/Matern 5/2 with a single correlation length, and the product form of the Matern 5/2 kernel.

Changes in GPU Version:

• Multi-Output GPs fit in parallel.

Other changes:

• Additional demo examples and documentation improvements.

v0.5.0

New release including GPU support and several other fixes and improvements.

Changes in this release:

• Support for GPU using the GaussianProcessGPU class.
  • Supports Squared Exponential and Matern Kernels
  • Supports basic polynomial mean functions which can be parsed from strings
  • Supports fixed, adaptive, and fit nuggets
  • Works with MultiOutputGP class, though fitting and predictions are done serially on the GPU.
• Pivoted Cholesky decomposition support for CPU GP fitting
• Modifications to MultiOutputGP fitting to allow GPs that fail to do so quietly and enables inspection of individual emulators based on those that failed.
• Improved demos and documentation.

v0.4.0

Minor update to update the repository name to put mogp on PyPI. Also adds a demo from the 25 September 2020 Excalibur workshop.

Changes in this release:

• Add excalibur_workshop_demo.py to the demos directory
• Change URLs to git repo
• mogp-emulator can now be installed via PyPI

v0.3.1

This is a micro bugfix release to fix an issue that occasionally came up when fitting emulators.

Changes in this release:

• Corrected a bug in the kernel distance metric computation and added tests to catch the situation.

v0.3.0

This is a major update, incorporating a complete rewrite and upgrade of the GP class and many of the other core library components. It is not backwards compatible with previous versions of the library and it has removed some pieces of the GP class, but it should improve flexibility going forward. A number of documentation additions have also been made to improve presentation and help users understand what is going on.

Specific improvements that have been made:

• Refactor of the GP class to externalize fitting. A GaussianProcess object now more closely represents the mathematical definition in that routines for estimation/fitting have been externalized. MCMC sampling for estimation has been removed for the moment in this process, though I anticipate being able to add it back in the future.
• Addition of mean functions. Mean functions can be added using a string formula in a similar manner to R. This optionally uses the patsy library for creating models (and in the future, design matrices), though patsy is not required to use the base mean function implementation.
• Basic support for prior distributions on hyperparameters has been added. This should be improved in future releases.
• Predictions can now include/exclude the nugget from the variance as desired.
• Documentation improvements and examples have been added.
• Numerous other bugfixes.

v0.2.0

Update master to v0.2.0. New major features in this release:

• Dimension reduction via the DimensionReduction class
• Fitting Emulators with MCMC sampling
• History matching via the HistoryMatching class
• Documentation now includes the MUCM toolkit

There are also several other minor changes, fixes, and improvements in this update, including improved prediction stability and addition of simple demo scripts.