NEWS.md
In dgpsi: Interface to 'dgpsi' for Deep and Linked Gaussian Process Emulations

dgpsi 2.5.0

Training times for DGP emulators are now approximately 30%-40% faster.
The computation of (D)GP predictions and Leave-One-Out (LOO) evaluations is now 6-7 times faster.
The nb_parallel argument has been removed from relevant functions, as multi-threading is now integrated by default.
A Vecchia approximation, implemented under the SI framework, is now available across all functions to support large-scale emulations.
Two new functions, get_thread_num() and set_thread_num(), allow users to inspect and adjust the number of threads used for multi-threaded computations.
A new function, set_vecchia(), enables users to easily add or remove the Vecchia approximation for GP, DGP, or linked (D)GP emulators.
Documentation now includes lifecycle status badges to highlight deprecated and newly introduced functions and arguments.
The default value of the nugget parameter in DGP emulators with likelihood layers has been adjusted from 1e-6 to 1e-4.
A Categorical likelihood option has been added to the dgp() function’s likelihood argument, enabling DGP-based classification.
An issue related to the LD_LIBRARY environment variable on Linux systems has been resolved via the init_py() function.
The lgp() function has been enhanced to accept connection information among emulators in the form of a data frame, streamlining linked emulation setup.
A new function, set_id(), allows users to assign unique IDs to emulators.
The predict() function has been updated to accommodate predictions from DGP classifiers.
The plot() function has been updated to generate validation plots for DGP classifiers (i.e., DGP emulators with categorical likelihoods) and linked emulators created by lgp() using the new data frame form for struc.
The summary() function has been redesigned to provide both summary tables and visualizations of structure and model specifications for (D)GP and linked (D)GP emulators.
A sample_size argument has been added to the validate() and plot() functions, allowing users to adjust the number of samples used for validation when the validation method is set to sampling.
combine() and set_linked_idx() are deprecated as of this version and will be removed in the next release. These two functions are no longer maintained. Please refer to the updated package documentation for alternative workflows.
The basic node functions kernel(), Hetero(), Poisson(), and NegBin(), along with the struc argument in the gp() and dgp() functions, have been removed as of this version. Customization of (D)GP specifications can be achieved by modifying the other arguments in gp() and dgp().
The draw() function has been updated for instances of the bundle class to allow drawing of design and evaluation plots of all emulators in a single figure.
The plot() function has been updated for linked emulators generated by lgp() using the new data frame form for struc.
The design() function has been redesigned to allow new specifications of the user-supplied method function.
The batch_size argument has been added to design() to enable locating multiple design points in a single iteration of the sequential design. This argument is compatible with all built-in method functions: alm(), mice(), and vigf().
The alm() and vigf() functions have been redesigned to support continuous search for the next design point or search from a discrete candidate set passed through the x_cand argument.
The alm(), mice(), and vigf() functions have been updated to output the locations of identified design points when a discrete candidate set is not supplied.
The pei() function has been removed from the package for re-engineering and will be added back in a future version.
The default of the refit argument in the update() function has been changed from FALSE to TRUE.
The write() function now allows light = TRUE for both GP emulators and bundles of GP emulators.
Two new functions, serialize() and deserialize(), have been added to allow users to export emulators to multi-session workers for parallel processing.
Additional vignettes are available, showcasing large-scale DGP emulation, DGP classification, and Bayesian optimization using (D)GP emulators.
Enhanced clarity and consistency across the documentation.
Improved examples and explanations in vignettes for better user guidance.

dgpsi 2.4.0

One can now use design() to implement sequential designs using f and a fixed candidate set passed to x_cand with y_cand = NULL.
The sizes of .pkl files written by write() are significantly reduced.
One can now set different kernel functions to nodes in different layers in a DGP emulator by passing a vector of kernel function names to name argument of dgp().
The default number of imputations B in dgp() and lgp() is changed to 10 for faster validations and predictions.
The default method for sequential designs in design() is changed to vigf().
A new argument new_wave is added to design() to allow users to resume sequential designs with or without a separate wave.
A bug in vigf() is fixed when object is an instance of the bundle class and batch_size is greater than one.
Static and dynamic pruning of DGP structures are implemented in prune() and design() (via the new arguments pruning and control) respectively.
Some redundant codes are removed from update() which makes design() slightly faster.
limits argument in design() is now required when x_cand is not supplied to avoid under-sampling using the limits inferred from the training data.
design() now supports f that produce NA as outputs. This is useful to prevent the sequential design from stopping due to errors or NA outputs from a simulator at some input locations identified by the sequential design process.
A bug is fixed in design() when x_cand is supplied and the input dimension is one.
alm(), mice(), pei(), and vigf() now accept separate candidate sets (even with different number of candidate points) via x_cand for bundle emulators.
A slot called id is added to instances of gp, dgp, lgp, and bundle classes to uniquely identify the emulators. id can also be passed to instances of gp, dgp,lgp, and bundle classes by the new id argument in gp(), dgp(), lgp(), and pack().
pack() can now accept a list of (D)GP emulators as the input.
The check_point argument is removed from design() and replaced by autosave.
Automatic saving of emulators during the sequential design is added to design() through the new argument autosave.
When a customized evaluation function is provided to design() via eval, the design information in previous waves will be retained as long as the previous waves of the sequential design also use customized evaluation functions. If different customized evaluation functions are supplied to design() in different waves, the trace plot of RMSEs produced by draw() will show RMSEs from different evaluation functions in different waves.
One can now link the same emulator multiple times in a chain via lgp() by setting different linking information for the emulator via set_linked_idx().
Updates of documentations and vignettes.

dgpsi 2.3.0

A bug from the underlying Python implementations is fixed when name = 'matern2.5' in gp() and dgp().
Thanks to @yyimingucl, a bug from the underlying Python implementations for the MICE sequential design criterion mice() is fixed.
An argument reset is added to update() and design() to reset hyperparameters of a (D)GP emulator to their initial values (that were specified when the emulator is initialized) after the input and output of the emulator are updated and before the emulator is refitted. This argument can be useful for sequential designs in cases where the hyperparameters of a (D)GP emulator get caught in suboptimal estimates. In such circumstances, one can set reset = TRUE to reinitialize the (D)GP emulator in some steps of the sequential designs as a strategy to escape the poor estimates.
The refitting of an emulator in the final step of a sequential design is no longer forced in design().
An argument type is added to plot() to allow users to draw OOS validation plots with testing data shown as a line instead of individual points when the emulator's input is one-dimensional and style = 1.
Thanks to @tjmckinley, an issue relating to libstdc++.so.6 on Linux machines when R is restarting after the installation of the package is fixed.
alm() and mice() can locate new design points for stochastic simulators with (D)GP or bundle emulators that can deal with stochastic outputs.
design() can be used to construct (D)GP or bundle emulators adaptively by utilizing multiple realizations from a stochastic simulator at the same design positions through the new argument reps when method = alm or method = mice.
A new slot called specs is added to the objects returned by gp() and dgp() that contains the key information of the kernel functions used in the constructions of GP and DGP emulators.
Due to a bug in the latest version of an underlying Python package, the emulators saved by write() in version 2.1.6 and 2.2.0 may not work properly with update() and design() when they are loaded back by read() in this version. This bug has been addressed in this version so emulators saved in this version would not have the compatibility issue in future version.
A new sequential design criterion, called the Variance of Improvement for Global Fit (VIGF), is added to the package with the function vigf().
The sampling from an existing candidate set x_cand in design() is changed from a random sampling to a conditioned Latin Hypercube sampling in clhs package.
The python environment is now automatically installed or invoked when a function from the package is executed. One does not need to run init_py() to activate the required python environment but init_py() is still useful to re-install and uninstall the underlying python environment. A verb argument is added to init_py() to switch on/off the trace information.

dgpsi 2.2.0

The efficiency and speed of imputations involved in the training and predictions of DGP emulators are significantly improved (achieving roughly 3x faster training and imputations) by utilizing blocked Gibbs sampling that imputes latent variables layer-wise rather than node-wise. The blocked Gibbs sampling is now the default method for DGP emulator inference and can be changed back to the old node-wise approach by setting blocked_gibbs = FALSE in dgp().
One can now optimize GP components that are contained in the same layer of a DGP emulator in parallel during the DGP emulator training, using multiple cores by setting the new argument cores in dgp(). This option is useful and can accelerate the training speed when the input dimension is moderately large (in which case there is a large number of GP components to be optimized) and the optimization of GP components is computationally expensive, e.g., when share = FALSE in which case input dimensions to individual GP components have different lengthscales.
Thanks to @tjmckinley, a bug in update() when the object is an instance of the dgp class (that has been trimmed by window()) is fixed.
Thanks to @tjmckinley, some R memory issues due to the underlying Python implementations are rectified.
set_seed() function is added to ensure reproducible results from the package.
A bug is fixed when candidate sets x_cand and y_cand are provided to design().
One can choose different color palettes using the new argument color in plot() when style = 2.
set_linked_idx() allows constructions of different (D)GP emulators (in terms of different connections to the feeding layers) from a same (D)GP emulator.

dgpsi 2.1.6

A bug is found in multi-core predictions in predict() when object is an instance of lgp class and x is a list. This bug has been fixed in this version.
Thanks to @tjmckinley, an issue (/usr/lib/x86_64-linux-gnu/libstdc++.so.6: version 'GLIBCXX_3.4.30' not found) encountered in Linux machines is fixed automatically during the execution of init_py().
gp() and dgp() allow users to specify the value of scale parameters and whether to estimate the parameters.
gp() and dgp() allow users to specify the bounds of lengthscales.
The jointly robust prior (Gu, 2019) is implemented as the default inference approach for GP emulators in gp().
The default value of lengthscale in gp() is changed from 0.2 to 0.1, and the default value for nugget in gp() is changed from 1e-6 to 1e-8 if nugget_est = FALSE.
One can now specify the number of GP nodes in each layer (except for the final layer) of a DGP emulator via the node argument in dgp().
Training data are now contained in the S3 classes gp and dgp.
The RMSEs (without the min-max normalization) of emulators are now contained in the S3 classes gp, dgp, and lgp after the execution of validate().
window() function is added to trim the traces and obtain new point estimates of DGP model parameters for predictions.
The min-max normalization can now be switched off in plot() by setting the value of min_max.
The default number of imputations B for dgp() is changed from 50 to 30 to better balance the uncertainty and the speed of DGP emulator predictions. A new function set_imp() is made available to change the number of imputations of a trained DGP emulator so one can either achieve faster predictions by further reducing the number of imputations, or account for more imputation uncertainties by increasing the number of imputations, without re-training the emulator.
The default number of imputations B for continue() is set to NULL, in which case the same number of imputations used in object will be applied.
nugget argument of dgp() now specifies the nugget values for GP nodes in different layers rather than GP nodes in the final layer.
The speed of predictions from DGP emulators with squared exponential kernels is significantly improved and is roughly 3x faster than the implementations in version 2.1.5.
The implementation of sequential designs (with two vignettes) of (D)GP emulators using different criterion is made available.
Thanks to @tjmckinley, an internal reordering issue in plot() is fixed.
init_py() now allows users to reinstall and uninstall the underlying Python environment.
A bug that occurs when a linked DGP emulator involves a DGP emulator with external inputs is fixed.
Intel SVML will now be installed with the Python environment automatically for Intel users for faster implementations.