In CollinErickson/CGGP: Composite Grid Gaussian Processes

# options(width = 500)
# options(fig_width="120pt")

knitr::opts_chunk$set(echo = TRUE)
knitr::opts_chunk$set(fig.width=12)
knitr::opts_chunk$set(fig.height=9)
library(magrittr)
library(ggplot2)
library(dplyr)

We want to compare our SGGP to other methods.

Experiment one

• SGGP

• laGP: Using all points, just a GP

• GPflow: Python, based on GPy, can scale to many points

• MRFA

• (5) functions as usual

• (4) n = 100, 250, 500, 750, 1000

e1 <- readRDS("./ExComp1_completed.rds")
e1df <- e1$outcleandf
e1$completed_runs %>% table

plyr::dlply(e1df %>% select(-start_time, -end_time, -run_number), "d")

ggplot(data=e1df, mapping=aes(n, RMSE)) + geom_point() + facet_grid(f ~ package, scales="free_y") + scale_y_log10()

ggplot(data=e1df, mapping=aes(n, score)) + geom_point() + facet_grid(f ~ package, scales="free_y")

ggplot(data=e1df, mapping=aes(n, CRPscore)) + geom_point() + facet_grid(f ~ package, scales="free_y") + scale_y_log10()

ggplot(data=e1df, mapping=aes(n, runtime)) + geom_point() + facet_grid(f ~ package, scales="free_y") + scale_y_log10()

CollinErickson/CGGP documentation built on Feb. 6, 2024, 2:24 a.m.