Releases · SMTorg/smt

@jbussemaker

Mixture of Experts improvements: (#282 thanks @jbussemaker, #283)
- add variance prediction API (ie. predict_variances()) which is enabled when variances_support option is set
- add MOESurrogateModel class which adapts MOE to the SurrogateModel interface
- allow selection of experts to be part of the mixture (see allow/deny options)
- MOE.AVAILABLE_EXPERTS lists all possible experts
- enabled_experts property of an MOE instance lists possible experts wrt derivatives/variances_support
  and allow/deny options.
Sampling Method interface refactoring: (#284 thanks @LDAP)
- create an intermediate ScaledSamplingMethod class to be the base class for sampling methods
  which generate samples in the [0, 1] hypercube
- allow future implementation of sampling methods generating samples direcly in the input space (i.e. within xlimits)
Use of Gower distance in kriging based mixed integer surrogate: (#289 thanks @raul-rufato )
- add use_gower_distance option to MixedIntegerSurrogate
- add gower correlation model to kriging based surrogate
- see MixedInteger notebook for usage
Improve kriging based surrogates with multistart method (#293 thanks @Paul-Saves )
- run several hyperparameter optimizations taking the best result
- number of optimization is controlled by n_start new option (default 10)
Update documentation for MOE and SamplingMethod (#285)
Fixes (#279, #281)

@anfelopera

Noise API changes for Kriging based surrogates (#276, #257 thanks @anfelopera):
- add a new tutorial notebook on how to deal with noise in SMT
- rename noise as noise0 option and is now a list of values
- add option use_het_noise to manage heteroscedastic noise,
- improve noise management for MFK (different noise by level),
- add option nugget to enable the handling of numerical instabilitily
- matern kernel documentation
Add predict_variance_derivatives API (#256 , #259 thanks @Paul-Saves)
- add spatial derivatives for Kriging based surrogates
- fix respect of parameters bounds in Kriging based surrogates
Notebooks updates (#262, #275 thanks @NatOnera, #277 thanks @Paul-Saves )
Kriging based surrogates refactoring (#261 thanks @anfelopera)
- inheritance changes: MFKPLS -> MFK, KPLSK, GEKPLS -> KPLS
- improve noise options consistency
- improve options validity checking
Code quality (#264, #267, #268 thanks @LDAP):
- use of abc metaclass to enforce developer API
- type hinting
- add 'build system' specification and requirements.txt for tests, setup cleanup

@repriem

add Marginal Gaussian Process surrogate models(#236, thanks @repriem)
add Matern kernels for kriging based surrogates (#236, thanks @repriem)
add gradient based optimization for hyperparameters in kriging based surrogates: new hyper_opt option to specify TNC Scipy gradient based optimizer, gradient-free Cobyla optimizer remains the default. (#236, thanks @repriem)
add MixedIntegerContext documentation (#234 )
fix bug in mixed_integer::unfold_with_enum_mask (#233 )

@Paul-Saves

Application: Mixed integer sampling methods and surrogates (#229)
- handling of categorical and integer variables in Kriging (#219, thanks @Paul-Saves)
- handling of categorical and integer variables in EGO optimizer (#220, thanks @Paul-Saves)
remove initial doe returned value from EGO optimize method (#224)
drop Python 2.7 (#215, #227)
fix MFK variance computation (#211)
fix MOE experts selection (#223)
fix MOE RMTS usage (#225)
fix QP as used in run_examples (#226)

@m-meliani

add MFKPLS application (#193, thanks @m-meliani )
add MFKPLSK application (#198, thanks @m-meliani )
add parallel EGO with qEI criterion (#202 , #199, #190, thanks @EmileRouxSMB and @rruusu )
add notebook tutorial for EGO method
fix kriging based surrogate bug (#200 )
fix full factorial sampling weights and clip options (#197 )
fix sklearn 0.22 cross_decomposition warning (#196 )
next releases > 0.5.x will likely drop Python 2.7 support

Releases: SMTorg/smt