Concepts¶
gpmp contains the numerical GP objects: gpmp.core.Model, covariance
functions, parameter-selection functions, diagnosis tools, posterior samplers,
and backend operations. gpmp-contrib adds ComputerExperiment,
ModelContainer, provided Matérn container classes, sequential strategies,
optimization and set-estimation criteria, benchmark problems, and reGP utilities.
Object map¶
ComputerExperimentDescribes an input domain and one or more callable outputs. Outputs can be objectives, constraints, or generic scalar functions. Constraint outputs may carry bounds that define feasibility.
ModelContainerStores one independent
gpmp.core.Modelper output. Each output entry stores the underlying model, selection criterion, raw parameter vectors,Paramobject, optimizer report, and resolved prior object when the model class uses REMAP priors.- Provided Matérn container classes
Configure Matérn GP models on top of
ModelContainer. They specify the mean form, covariance function, initial-guess procedure, selection criterion, andParamconstruction.SequentialPredictionStores observations and a model container. It updates model parameters, computes predictions, and draws conditional simulations from the current model.
- Sequential strategy classes
Add a candidate set or particle set and a rule for selecting the next evaluation. The implemented rules cover expected improvement, excursion sets, set inversion, and particle-based searches.
Common notation¶
The guide uses xi for observation points and zi for observed values.
Mathematically, write
For a scalar output, the provided Matérn classes build a Gaussian process
where \(Z_0\) is centered and has covariance
\(K_\theta(x,x')\). A multi-output ModelContainer stores one
independent scalar model per output. Cross-output covariance is not modeled by
the provided container classes.
After conditioning on observations, predictions at a point \(x\) are Gaussian:
The arrays returned by predict are these posterior means and variances,
stacked over prediction points and outputs.
Relation to gpmp¶
The GP model is the standard kriging model used in spatial statistics and
computer experiments. Background on covariance modeling and kriging is given by
[2, 7]. In gpmp-contrib, the
model formulas are not reimplemented. Prediction, likelihood evaluation, REML,
REMAP, and posterior parameter sampling are delegated to gpmp.
gpmp-contrib decides how the model is assembled and how state is stored:
which covariance class is used, which parameter-selection criterion is called,
how optimizer starts are chosen, and how user-readable parameter objects are
created.
When to use gpmp-contrib¶
Use gpmp-contrib when the task involves a computer experiment, a
multi-output model container, a sequential design, a provided benchmark problem,
or the reGP procedure. Use gpmp directly when implementing a new covariance
function, selection criterion, MCMC kernel, backend operation, or diagnostic.