The hardware and bandwidth for this mirror is donated by METANET, the Webhosting and Full Service-Cloud Provider.
If you wish to report a bug, or if you are interested in having us mirror your free-software or open-source project, please feel free to contact us at mirror[@]metanet.ch.
design() to implement sequential designs using f and a fixed candidate set passed to x_cand with y_cand = NULL..pkl files written by write() are significantly reduced.name argument of dgp().B in dgp() and lgp() is changed to 10 for faster validations and predictions.design() is changed to vigf().new_wave is added to design() to allow users to resume sequential designs with or without a separate wave.vigf() is fixed when object is an instance of the bundle class and batch_size is greater than one.prune() and design() (via the new arguments pruning and control) respectively.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.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.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.check_point argument is removed from design() and replaced by autosave.design() through the new argument autosave.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.lgp() by setting different linking information for the emulator via set_linked_idx().name = 'matern2.5' in gp() and dgp().mice() is fixed.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.design().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.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.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.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.vigf().x_cand in design() is changed from a random sampling to a conditioned Latin Hypercube sampling in clhs package.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.blocked_gibbs = FALSE in dgp().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.update() when the object is an instance of the dgp class (that has been trimmed by window()) is fixed.set_seed() function is added to ensure reproducible results from the package.x_cand and y_cand are provided to design().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.predict() when object is an instance of lgp class and x is a list. This bug has been fixed in this version./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.gp().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.node argument in dgp().gp and dgp.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.plot() by setting the value of min_max.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.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.2.1.5.plot() is fixed.init_py() now allows users to reinstall and uninstall the underlying Python environment.Intel SVML will now be installed with the Python environment automatically for Intel users for faster implementations.dgpsi v2.1.5.These binaries (installable software) and packages are in development.
They may not be fully stable and should be used with caution. We make no claims about them.