In berndbischl/mlrMBO: Bayesian Optimization and Model-Based Optimization of Expensive Black-Box Functions

Introductory example

Info: this guide gives an overview of the typical optimization workflow with mlrMBO for those familiar with model based optimization. For a much more detailed introduction see the next chapter.

In this example we aim to maximize a one dimensional mixed cosine function using model-based optimization. Instead of writing this function by hand, we make use of the smoof package, which offers a lot of common single objective optimization functions.

library(mlrMBO)
obj.fun = makeCosineMixtureFunction(1)
plot(obj.fun)

We decide to use kriging as our surrogate model and to do 10 sequential optimization steps. Furthermore we use Expected Improvement (EI) as the infill criterion.

As a last step we have to generate an initial design on which we evaluate our model in the beginning. We use ParamHelpers::generateDesign to generate 10 points in a latin hypercube design.

learner = makeLearner("regr.km", predict.type = "se", covtype = "matern3_2", 
  control = list(trace = FALSE))
control = makeMBOControl()
control = setMBOControlTermination(control, iters = 10)
control = setMBOControlInfill(control, crit = "ei")
design = generateDesign(n = 10, par.set = getParamSet(obj.fun))

Finally we start the optimization process and print the result object.

result = mbo(obj.fun, design = design, learner = learner, control = control, 
             show.info = TRUE)
print(result)

Example run

There is also the function exampleRun, which is useful to figure out how MBOworks and to visualize the results.

ex = exampleRun(obj.fun, learner = learner, control = control, show.info = FALSE)

print(ex)
plotExampleRun(ex, iters = c(1L, 3L, 10L))

Or alternatively for a two dimensional function:

obj.fun2 = makeCosineMixtureFunction(2L)
plot(obj.fun2)
ex2 = exampleRun(obj.fun2, learner = learner, control = control, show.info = FALSE)

print(ex2)
plotExampleRun(ex2, iters = c(1L, 3L, 10L))

berndbischl/mlrMBO documentation built on Oct. 11, 2022, 1:44 p.m.