README.md
In paradox: Define and Work with Parameter Spaces for Complex Algorithms

paradox

Package website: release | dev

Universal Parameter Space Description and Tools.

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

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)
#> <Design> with 3 rows:
#>    z         x  flag methods
#> 1: 1  7.660348 FALSE       b
#> 2: 3  8.809346 FALSE       c
#> 3: 2 -9.088870 FALSE       b

Generate LHS Design:

requireNamespace("lhs")
#> Loading required namespace: lhs
generate_design_lhs(pset, 3)
#> <Design> with 3 rows:
#>    z         x  flag methods
#> 1: 1 -3.984673  TRUE       b
#> 2: 2  7.938035 FALSE       a
#> 3: 3  1.969783  TRUE       c

Generate Grid Design:

generate_design_grid(pset, resolution = 2)
#> <Design> with 24 rows:
#>     z   x  flag methods
#>  1: 1 -10  TRUE       a
#>  2: 1 -10  TRUE       b
#>  3: 1 -10  TRUE       c
#>  4: 1 -10 FALSE       a
#>  5: 1 -10 FALSE       b
#>  6: 1 -10 FALSE       c
#>  7: 1  10  TRUE       a
#>  [ reached getOption("max.print") -- omitted 18 rows ]

Properties of the parameters within the ParamSet:

pset$ids()
#> [1] "z"       "x"       "flag"    "methods"
pset$levels
#> $z
#> NULL
#> 
#> $x
#> NULL
#> 
#> $flag
#> [1]  TRUE FALSE
#> 
#> $methods
#> [1] "a" "b" "c"
pset$nlevels
#>       z       x    flag methods 
#>       3     Inf       2       3
pset$is_number
#>       z       x    flag methods 
#>    TRUE    TRUE   FALSE   FALSE
pset$lower
#>       z       x    flag methods 
#>       1     -10      NA      NA
pset$upper
#>       z       x    flag methods 
#>       3      10      NA      NA

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$test(list(z = 1, x = 1))
#> [1] TRUE
pset$test(list(z = -1, x = 1))
#> [1] FALSE
pset$check(list(z = -1, x = 1))
#> [1] "z: Element 1 is not >= 0.5"
pset$assert(list(z = -1, x = 1))
#> Error in pset$assert(list(z = -1, x = 1)): Assertion on 'list(z = -1, x = 1)' failed: z: Element 1 is not >= 0.5.

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
#> [[1]]
#> [[1]]$z
#> [1] 0.125
#> 
#> [[1]]$x
#> [1] 0.6985067
#> 
#> 
#> [[2]]
#> [[2]]$z
#> [1] 0.5
#> 
#> [[2]]$x
#> [1] 0.5795772

Further documentation can be found in the in-depth tutorial