Flexible Bayesian Optimization.

mlr3mbo

Package website: release | dev

A new R6 and much more modular implementation for single- and multi-objective Bayesian Optimization.

Get Started

The best entry point to get familiar with mlr3mbo is provided via the Bayesian Optimization chapter in the mlr3book.

Design

mlr3mbo is built modular relying on the following R6 classes:

• Surrogate: Surrogate Model
• AcqFunction: Acquisition Function
• AcqOptimizer: Acquisition Function Optimizer

Based on these, Bayesian Optimization (BO) loops can be written, see, e.g., bayesopt_ego for sequential single-objective BO.

mlr3mbo also provides an OptimizerMbo class behaving like any other Optimizer from the bbotk package as well as a TunerMbo class behaving like any other Tuner from the mlr3tuning package.

mlr3mbo uses sensible defaults for the Surrogate, AcqFunction, AcqOptimizer, and even the loop_function. See ?mbo_defaults for more details.

Simple Optimization Example

Minimize the two-dimensional Branin function via sequential BO using a GP as surrogate and EI as acquisition function optimized via a local serch:

library(bbotk)
library(mlr3mbo)
library(mlr3learners)
set.seed(1)

fun = function(xdt) {
  y = branin(xdt[["x1"]], xdt[["x2"]])
  data.table(y = y)
}

domain = ps(
  x1 = p_dbl(-5, 10),
  x2 = p_dbl(0, 15)
)

codomain = ps(
  y = p_dbl(tags = "minimize")
)

objective = ObjectiveRFunDt$new(
  fun = fun,
  domain = domain,
  codomain = codomain
)

instance = oi(
  objective = objective,
  terminator = trm("evals", n_evals = 25)
)

surrogate = srlrn(lrn("regr.km", control = list(trace = FALSE)))

acq_function = acqf("ei")

acq_optimizer = acqo(
  opt("local_search", n_initial_points = 10, initial_random_sample_size = 1000, neighbors_per_point = 10),
  terminator = trm("evals", n_evals = 2000)
)

optimizer = opt("mbo",
  loop_function = bayesopt_ego,
  surrogate = surrogate,
  acq_function = acq_function,
  acq_optimizer = acq_optimizer
)

optimizer$optimize(instance)

##          x1       x2  x_domain        y
##       <num>    <num>    <list>    <num>
## 1: 3.104516 2.396279 <list[2]> 0.412985

We can quickly visualize the contours of the objective function (on log scale) as well as the sampling behavior of our BO run (lighter blue colours indicating points that were evaluated in later stages of the optimization process; the first batch is given by the initial design).

library(ggplot2)
grid = generate_design_grid(instance$search_space, resolution = 1000L)$data
grid[, y := branin(x1 = x1, x2 = x2)]

ggplot(aes(x = x1, y = x2, z = log(y)), data = grid) +
  geom_contour(colour = "black") +
  geom_point(aes(x = x1, y = x2, colour = batch_nr), data = instance$archive$data) +
  labs(x =  expression(x[1]), y = expression(x[2])) +
  theme_minimal() +
  theme(legend.position = "bottom")

Note that you can also use bb_optimize as a shorthand instead of constructing an optimization instance.

Simple Tuning Example

library(mlr3)
library(mlr3learners)
library(mlr3tuning)
library(mlr3mbo)
set.seed(1)

task = tsk("pima")

learner = lrn("classif.rpart", cp = to_tune(lower = 1e-04, upper = 1, logscale = TRUE))

instance = tune(
  tuner = tnr("mbo"),
  task = task,
  learner = learner,
  resampling = rsmp("holdout"),
  measure = msr("classif.ce"),
  term_evals = 10)

instance$result

##           cp learner_param_vals  x_domain classif.ce
##        <num>             <list>    <list>      <num>
## 1: -6.188733          <list[2]> <list[1]>  0.2382812