Mixed Integer Evolution Strategies

Evolutionary black box optimization algorithms building on the 'bbotk' package. 'miesmuschel' offers both ready-to-use optimization algorithms, as well as their fundamental building blocks that can be used to manually construct specialized optimization loops. The Mixed Integer Evolution Strategies as described by Li et al. (2013) can be implemented, as well as the multi-objective optimization algorithms NSGA-II by Deb, Pratap, Agarwal, and Meyarivan (2002) .

miesmuschel: Mixed Integer Evolution Strategies

What does it do?

miesmuschel provides evolutionary black box optimization algorithms, building on the bbotk package. miesmuschel offers both ready-to-use optimization algorithms, as well as their fundamental building blocks that can be used to manually construct specialized optimization loops.

What does it do? (poem version)

(also giving a hint on how to pronounce "miesmuschel")

The R software package miesmuschel
Offers opt-algorithms, so crucial
bbotk is its base, it's a powerful tool
For optimization, it's truly no fool

Ready-made or D-I-Y, the choice is yours
With miesmuschel, your options are wide open doors
So when you do optimize, just give it a try
With miesmuschel, surely, success is nigh!

Project Status

Although miesmuschel is currently still evolving, it can already be used for optimization. All exported functions are thoroughly documented.

Installation

Install the github version, using remotes:

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

Some Code to Get Started

Preparation

library("bbotk")
library("paradox")
lgr::threshold("warn")

objective <- ObjectiveRFun$new(
  fun = function(xs) {
    z <- exp(-xs$x^2 - xs$y^2) + 2 * exp(-(2 - xs$x)^2 - (2 - xs$y)^2)
    list(Obj = z)
  },
  domain = ps(x = p_dbl(-2, 4), y = p_dbl(-2, 4)),
  codomain = ps(Obj = p_dbl(tags = "maximize"))
)

Using bbotk::Optimizer Object

This is the recommended way of using miesmuschel.

# Get a new OptimInstance
oi <- OptimInstanceSingleCrit$new(objective,
  terminator = trm("evals", n_evals = 100)
)

library("miesmuschel")

# Get operators
op.m <- mut("gauss", sdev = 0.1)
op.r <- rec("xounif", p = .3)
op.parent <- sel("random")
op.survival <- sel("best")

# Create OptimizerMies object
mies <- opt("mies", mutator = op.m, recombinator = op.r,
  parent_selector = op.parent, survival_selector = op.survival,
  mu = 3, lambda = 2)

# mies$optimize performs MIES optimization and returns the optimum
mies$optimize(oi)
#>           x        y  x_domain      Obj
#> 1: 1.935055 1.973867 <list[2]> 1.990703

Using mies_* Functions Directly

This gives more flexibility when designing ES algorithms, but it is also more verbose and error-prone.

# Get a new OptimInstance
oi <- OptimInstanceSingleCrit$new(objective,
  terminator = trm("evals", n_evals = 100)
)

library("miesmuschel")

# Get operators
op.m <- mut("gauss", sdev = 0.1)
op.r <- rec("xounif", p = .3)
op.parent <- sel("random")
op.survival <- sel("best")

# Prime operators
mies_prime_operators(list(op.m), list(op.r), list(op.parent, op.survival),
  search_space = oi$search_space)

# Sample first generation
mies_init_population(oi, 3)

# This is the first generation
oi$archive$data[, .(x, y, Obj, dob, eol)]
#>             x          y        Obj dob eol
#> 1:  3.8516312  1.2386550 0.03633278   1  NA
#> 2: -1.6478480 -0.9080712 0.02901343   1  NA
#> 3: -0.4215587  0.8250017 0.42529339   1  NA

# Select parents, recombine, mutate
offspring <- mies_generate_offspring(oi, 2, op.parent, op.m, op.r)

# This is the first offspring population
offspring
#>            x           y
#> 1:  2.762783 -0.24885684
#> 2: -1.439780 -0.05699817

# Evaluate offspring (and append to oi archive)
mies_evaluate_offspring(oi, offspring)

# State of the archive now: Second generation has `dob` == 2 
oi$archive$data[, .(x, y, Obj, dob, eol)]
#>             x           y         Obj dob eol
#> 1:  3.8516312  1.23865501 0.036332776   1  NA
#> 2: -1.6478480 -0.90807120 0.029013430   1  NA
#> 3: -0.4215587  0.82500173 0.425293387   1  NA
#> 4:  2.7627829 -0.24885684 0.007566604   2  NA
#> 5: -1.4397798 -0.05699817 0.125404228   2  NA

# Selecto for survival
mies_survival_plus(oi, 3, op.survival)

# Survivors have `eol` NA, two individuals 'died' in generation 2
oi$archive$data[, .(x, y, Obj, dob, eol)]
#>             x           y         Obj dob eol
#> 1:  3.8516312  1.23865501 0.036332776   1  NA
#> 2: -1.6478480 -0.90807120 0.029013430   1   2
#> 3: -0.4215587  0.82500173 0.425293387   1  NA
#> 4:  2.7627829 -0.24885684 0.007566604   2   2
#> 5: -1.4397798 -0.05699817 0.125404228   2  NA

# Perform MIES loop until terminated. This gives an expected `terminated` error
repeat {
  offspring <- mies_generate_offspring(oi, 2, op.parent, op.m, op.r)
  mies_evaluate_offspring(oi, offspring)
  mies_survival_plus(oi, 3, op.survival)
}

# Best result:
oi$archive$data[which.max(Obj)]
#>           x        y dob eol      Obj  x_domain           timestamp batch_nr
#> 1: 2.021887 2.164051  46  NA 1.946115 <list[2]> 2021-02-15 01:56:23       46