lgpr2

Differences compared to the original lgpr package

• lgpr2 uses CmdstanR as backend, whereas lgpr uses rstan
• lgpr2 generates and compiles new Stan code each time a model is specified, whereas lgpr uses precompiled Stan code
• lgpr2 uses approximate GPs (basis function implementation) whereas lgpr uses exact ones (covariance matrix implementation)
• lgpr2 is not on CRAN
• lgpr2 is a lot faster and scales to bigger data
• lgpr2 has less features (for example, currently only supports Gaussian observation model)
• Stan code generated by lgpr2 is more readable and can be easily customized
• lgpr2 contains the projection predictive method additional model reduction technique

Requirements

• CmdStanR

Installation

devtools::install_github('jtimonen/lgpr2')

Reference

Juho Timonen & Harri Lähdesmäki, Scalable mixed-domain Gaussian process modeling and model reduction for longitudinal data, arXiv (2024)