This post describes the R version of Python’s GPopt (https://docs.techtonique.net/GPopt/GPopt.html), a package for ‘Bayesian’ optimization of black-box functions and machine learning hyperparameter tuning, using Gaussian Process Regression and other conformalized surrogates. The package is available on GitHub, and through the R universe.

Keep in mind that this package is for Machine Learning hyperparameter tuning: the global minimum won’t always be found, but this isn’t an issue, since it means you aren’t overfitting the training set.

It’s ported the same way as nnetsauce for R was: with uv to create an isolated Python virtual environment containing the Python GPopt package, and reticulate to call into it from R. Every function in this R package is a thin wrapper that returns the underlying Python object; the general rule is: object accesses with .’s in Python are replaced by $’s in R.

See this post for the technique: Finally figured out a way to port python packages to R using uv and reticulate.

Install

1. Create a Python virtual environment with uv

# pip install uv # if necessary
uv venv venv
source venv/bin/activate      # on Windows: venv\Scripts\activate
uv pip install pip GPopt

Keep track of where venv/ lives – you’ll pass its path as venv_path to every function in this package.

2. Install the R package

install.packages("remotes")
remotes::install_github("Techtonique/GPopt_r") 

reticulate will be installed automatically as a dependency.

Examples

Minimizing the Branin function

This is a standard test function for optimization algorithms. GPOpt is more suitable for expensive black-box functions, but this is a good example to illustrate the usage of the package.

library(GPopt)

branin <- function(x) {
  x1 <- x[1]; x2 <- x[2]
  term1 <- (x2 - (5.1 * x1^2) / (4 * pi^2) + (5 * x1) / pi - 6)^2
  term2 <- 10 * (1 - 1 / (8 * pi)) * cos(x1)
  term1 + term2 + 10
}

opt <- GPOpt(
  lower_bound = c(-5, 0),
  upper_bound = c(10, 15),
  objective_func = branin,
  n_init = 10,
  n_iter = 40,
  venv_path = "./venv"
)

opt$optimize(verbose = 1L)
print(opt$x_min)  # best parameters
print(opt$y_min)  # best objective value

Tuning a scikit-learn model’s hyperparameters

library(GPopt)

sklearn <- get_sklearn(venv_path = "./venv")
RandomForestClassifier <- sklearn$ensemble$RandomForestClassifier

X <- as.matrix(iris[, 1:4])
y <- as.integer(iris$Species) - 1L

mlopt <- MLOptimizer(scoring = "accuracy", cv = 5, venv_path = "./venv")

param_config <- list(
  n_estimators = list(bounds = c(10, 300), dtype = "int"),
  max_depth    = list(bounds = c(1, 20),   dtype = "int")
)

mlopt$optimize(
  X_train = X, y_train = y,
  estimator_class = RandomForestClassifier(),
  param_config = param_config,
  verbose = 1L
)

print(mlopt$get_best_parameters())
print(mlopt$get_best_score())

Bayesian optimization with early stopping

library(GPopt)

opt <- BOstopping(
  f = branin,
  bounds = rbind(c(-5, 10), c(0, 15)),
  venv_path = "./venv"
)
result <- opt$optimize(n_iter = 100L)

Using a custom (conformalized) surrogate model

library(GPopt)

sklearn <- get_sklearn(venv_path = "./venv")
ns <- get_nnetsauce(venv_path = "./venv")

opt <- GPOpt(
  lower_bound = c(-5, 0),
  upper_bound = c(10, 15),
  objective_func = branin,
  acquisition="ucb",
  method="splitconformal",
  surrogate_obj = ns$PredictionInterval(sklearn$ensemble$RandomForestRegressor()),
  venv_path = "./venv"
)
opt$optimize(verbose = 1L)