| Title: | Localized Ensemble of Approximate Gaussian Processes |
|---|---|
| Description: | An emulator designed for rapid sequential emulation (e.g., Markov chain Monte Carlo applications). Works via extension of the 'laGP' approach by Gramacy and Apley (2015 <doi:10.1080/10618600.2014.914442>). Details are given in Rumsey et al. (2023 <doi:10.1002/sta4.576>). |
| Authors: | Kellin Rumsey [aut, cre] |
| Maintainer: | Kellin Rumsey <[email protected]> |
| License: | GPL (>= 3) |
| Version: | 1.0.0 |
| Built: | 2025-03-01 03:50:18 UTC |
| Source: | https://github.com/cran/leapgp |
Function to train or initialize a leapGP model, as described in Rumsey et al. (2023).
leapGP( X, y, M0 = ceiling(sqrt(length(y))), rho = NA, scale = FALSE, n = ceiling(sqrt(length(y))), start = NA, verbose = FALSE, justdoit = FALSE, ... )leapGP( X, y, M0 = ceiling(sqrt(length(y))), rho = NA, scale = FALSE, n = ceiling(sqrt(length(y))), start = NA, verbose = FALSE, justdoit = FALSE, ... )
X |
a matrix of training locations (1 row for each training instance) |
y |
a vector of training responses ( |
M0 |
the number of prediction hubs desired. Defaults to |
rho |
(optional). The parameter controlling time-accuracy tradeoff. Can also be specified during prediction. |
scale |
logical. Do we want the scale parameter to be returned for predictions? If TRUE,
the matrix |
n |
local neighborhood size (for laGP) |
start |
number of starting points for neighborhood (between 6 and n inclusive) |
verbose |
logical. Should status be printed? Deault is FALSE |
justdoit |
logical. Force leapGP to run using specified parameters (may take a long time and/or cause R to crash). |
... |
optional arguments to be passed to |
The leapGP is extends the laGP framework of Gramacy & Apley (2015). The methods are equivalent for rho=1,
but leapGP trades memory for speed when rho < 1. The method is described in Rumsey et al. (2023) where they demonstrate
that leapGP is faster than laGP for sequential predictions and is also generally more accurate for some settings of rho.
an object of class leapGP with fields X, y, and hubs. Also returns scale parameter if scale=TRUE
Gramacy, R. B., & Apley, D. W. (2015). Local Gaussian process approximation for large computer experiments. Journal of Computational and Graphical Statistics, 24(2), 561-578.
Rumsey, K. N., Huerta, G., & Derek Tucker, J. (2023). A localized ensemble of approximate Gaussian processes for fast sequential emulation. Stat, 12(1), e576.
# Generate data f <- function(x){ 1.3356*(1.5*(1-x[1]) + exp(2*x[1] - 1)*sin(3*pi*(x[1] - 0.6)^2) + exp(3*(x[2]-0.5))*sin(4*pi*(x[2] - 0.9)^2)) } X <- matrix(runif(200), ncol=2) y <- apply(X, 1, f) # Generate data for prediction Xtest <- matrix(runif(200), ncol=2) ytest <- apply(Xtest, 1, f) # Train initial model mod <- leapGP(X, y, M0 = 30) # Make sequential predictions pred <- rep(NA, 100) for(i in 1:100){ mod <- predict_leapGP(mod, matrix(Xtest[i,], nrow=1), rho=0.9) pred[i] <- mod$mean }# Generate data f <- function(x){ 1.3356*(1.5*(1-x[1]) + exp(2*x[1] - 1)*sin(3*pi*(x[1] - 0.6)^2) + exp(3*(x[2]-0.5))*sin(4*pi*(x[2] - 0.9)^2)) } X <- matrix(runif(200), ncol=2) y <- apply(X, 1, f) # Generate data for prediction Xtest <- matrix(runif(200), ncol=2) ytest <- apply(Xtest, 1, f) # Train initial model mod <- leapGP(X, y, M0 = 30) # Make sequential predictions pred <- rep(NA, 100) for(i in 1:100){ mod <- predict_leapGP(mod, matrix(Xtest[i,], nrow=1), rho=0.9) pred[i] <- mod$mean }
Predict method for an object of class leapGP. Returns a (possibly modified) leapGP object as well as a prediction (with uncertainty, if requested).
predict_leapGP( object, newdata, rho = 0.95, scale = FALSE, n = ceiling(sqrt(length(y))), start = NA, M_max = Inf, ... )predict_leapGP( object, newdata, rho = 0.95, scale = FALSE, n = ceiling(sqrt(length(y))), start = NA, M_max = Inf, ... )
object |
An object of class |
newdata |
New data |
rho |
parameter controlling time-accuracy tradeoff (default is |
scale |
logical. Do we want the scale parameter to be returned for predictions? If TRUE,
the matrix |
n |
local neighborhood size |
start |
number of starting points for neighborhood (between 6 and n inclusive) |
M_max |
the maximum number of hubs allowed (used to upper bound the run time) |
... |
optional arguments to be passed to |
The leapGP is extends the laGP framework of Gramacy & Apley (2015). The methods are equivalent for rho=1,
but leapGP trades memory for speed when rho < 1. The method is described in Rumsey et al. (2023) where they demonstrate
that leapGP is faster than laGP for sequential predictions and is also generally more accurate for some settings of rho.
A list containing values mean, hubs X and y. If scale=TRUE the list also contains field sd.
Gramacy, R. B., & Apley, D. W. (2015). Local Gaussian process approximation for large computer experiments. Journal of Computational and Graphical Statistics, 24(2), 561-578.
Rumsey, K. N., Huerta, G., & Derek Tucker, J. (2023). A localized ensemble of approximate Gaussian processes for fast sequential emulation. Stat, 12(1), e576.
# Generate data f <- function(x){ 1.3356*(1.5*(1-x[1]) + exp(2*x[1] - 1)*sin(3*pi*(x[1] - 0.6)^2) + exp(3*(x[2]-0.5))*sin(4*pi*(x[2] - 0.9)^2)) } X <- matrix(runif(200), ncol=2) y <- apply(X, 1, f) # Generate data for prediction Xtest <- matrix(runif(200), ncol=2) ytest <- apply(Xtest, 1, f) # Train initial model mod <- leapGP(X, y, M0 = 30) # Make sequential predictions pred <- rep(NA, 100) for(i in 1:100){ mod <- predict_leapGP(mod, matrix(Xtest[i,], nrow=1), rho=0.9) pred[i] <- mod$mean }# Generate data f <- function(x){ 1.3356*(1.5*(1-x[1]) + exp(2*x[1] - 1)*sin(3*pi*(x[1] - 0.6)^2) + exp(3*(x[2]-0.5))*sin(4*pi*(x[2] - 0.9)^2)) } X <- matrix(runif(200), ncol=2) y <- apply(X, 1, f) # Generate data for prediction Xtest <- matrix(runif(200), ncol=2) ytest <- apply(Xtest, 1, f) # Train initial model mod <- leapGP(X, y, M0 = 30) # Make sequential predictions pred <- rep(NA, 100) for(i in 1:100){ mod <- predict_leapGP(mod, matrix(Xtest[i,], nrow=1), rho=0.9) pred[i] <- mod$mean }