Limited parallel fitting and acquisition searches to two worker processes.
Added random and space-filling threshold-accepting options for the sequence portion of an initial design. The existing simulated-annealing method remains the default for backward compatibility.
Added C++ incremental scoring for the threshold-accepting search, including bounded initialization, threshold calibration, best-result retention, and detailed diagnostics.
Connected the new sequence choices to the full initial-design and from-scratch Bayesian-optimization workflows.
Added examples and regression tests for method selection, input validation, reproducibility, finite permutation spaces, and quantitative-design preservation.
Rewrote the Bayesian optimization help pages and examples to explain function choice, arguments, workflow steps, and returned objects more directly.
Added a pkgdown website with a getting-started guide, an optimization workflow, a reproducible covariance-engine benchmark, and a curated function reference.
Added package-level citation metadata for both the software and the associated methodology paper.
Added magp_bayes_optimize_from_scratch() to connect
simulated-annealing initial design, initial objective evaluation, and
sequential MaGP Bayesian optimization in one call.
Added reproducible handling of stochastic objective functions while preserving the caller’s random-number state.
Included the complete initial design and its responses in the optimization result.
Added early validation for design, fitting, acquisition, and objective controls before the initial-design search begins.
Added expected improvement for both two-dimensional and full-mapping models.
Added mixed acquisition search across quantitative bounds and sequence permutations, with multi-start and optional parallel execution.
Added a sequential Bayesian optimization interface for expensive objective functions.
Added deterministic sequence sampling for cases where complete permutation enumeration would be too large.
Added duplicate-input protection, early stopping, complete search history, and final-model refitting.
Added multi-start optimization to magp2d_fit() and
magpfull_fit().
Added optional local parallel execution through portable socket clusters.
Made seeded sequential and parallel fits use the same parameter starts.
Added start-level objective values, convergence results, warnings, and errors to each fitted object.
Completed the quantitative-sequence initial-design workflow.
Added magp_quantitative_criterion() and
magp_quantitative_design() for maximin-style Latin
hypercubes.
Added magp_joint_criterion() to evaluate the
combined separation of quantitative and sequence inputs.
Added magp_initial_design() to construct and align
both design portions while preserving their individual
structures.
Added reproducible simulated-annealing searches and validation across several run sizes and component counts.
Added tools for constructing the sequence portion of a quantitative-sequence initial design.
Added magp_sequence_criterion() to assess pair
balance and Hamming-distance space filling in a sequence
design.
Added magp_sequence_design() to optimize the
sequence portion with reproducible simulated annealing.
Added reproducible simulated-annealing search with configurable criterion weights and search settings.
predict() interface, including
plug-in predictive uncertainty, for both fitted model classes.Rcpp.tau = 0.001.magp.predict() methods for both fitted model
classes.Rcpp for covariance, gradient, and
cross-covariance calculations.[0, 1].