In mlr-org/paradox: Define and Work with Parameter Spaces for Complex Algorithms

paradox

Package website: release | dev

Universal Parameter Space Description and Tools.

set.seed(123)
knitr::opts_chunk$set(cache = FALSE, collapse = TRUE, comment = "#>")
options(datatable.print.class = FALSE, datatable.print.keys = FALSE)
library(paradox)

Installation

remotes::install_github("mlr-org/paradox")

Usage

Create a simple ParamSet using all supported Parameter Types:

• integer numbers ("int")
• real-valued numbers ("dbl")
• truth values TRUE or FALSE ("lgl")
• categorical values from a set of possible strings ("fct")
• further types are only possible by using transformations.

pset = ps(
  z = p_int(lower = 1, upper = 3),
  x = p_dbl(lower = -10, upper = 10),
  flag = p_lgl(),
  methods = p_fct(c("a","b","c"))
)

Draw random samples / create random design:

generate_design_random(pset, 3)

Generate LHS Design:

requireNamespace("lhs")
generate_design_lhs(pset, 3)

Generate Grid Design:

generate_design_grid(pset, resolution = 2)

Properties of the parameters within the ParamSet:

pset$ids()
pset$levels
pset$nlevels
pset$is_number
pset$lower
pset$upper

Parameter Checks

Check that a parameter satisfies all conditions of a ParamSet, using $test() (returns FALSE on mismatch), $check() (returns error description on mismatch), and $assert() (throws error on mismatch):

pset$test(list(z = 1, x = 1))
pset$test(list(z = -1, x = 1))
pset$check(list(z = -1, x = 1))
pset$assert(list(z = -1, x = 1))

Transformations

Transformations are functions with a fixed signature.

• x A named list of parameter values
• param_set the ParamSet used to create the design

Transformations can be used to change the distributions of sampled parameters. For example, to sample values between $2^-3$ and $2^3$ in a $log_2$-uniform distribution, one can sample uniformly between -3 and 3 and exponentiate the random value inside the transformation. Alternatively, logscale = TRUE can be set; in this case, lower and upper represent the values after the transformation.

pset = ps(
  z = p_int(lower = -3, upper = 3),
  x = p_dbl(lower = 2^-3, upper = 2^3, logscale = TRUE)
)
pset$extra_trafo = function(x, param_set) {
  x$z = 2^x$z
  return(x)
}
pset_smplr = SamplerUnif$new(pset)
x = pset_smplr$sample(2)
xst = x$transpose()
xst

Further documentation can be found in the in-depth tutorial

mlr-org/paradox documentation built on Feb. 1, 2025, 9:42 p.m.