Model-Based Optimization for 'mlr3' Through 'mlrMBO'.
mlrintermbo: mlrMBO mlr3 Interface
What Is This?
Currently, it is not possible to use mlrMBO for tuning mlr3 and related packages directly, because of some disagreements between S3 (as used in mlrMBO) and R6 (used in mlr3). mlr3mbo exists, but it is not yet as mature and feature-rich as mlrMBO. mlrintermbo
provides the necessary interface to make mlrMBO accessible for mlr3.
To use mlrintermbo
, one should NOT load mlrMBO
as a library in the current R session. Instead, mlrintermbo
will run mlrMBO on a different background R session to keep it sectioned off from the main process. Just load the tuner (for tuning mlr3 "Learners") or optimizer (for tuning bbotk "Objectives"):
library("mlrintermbo")
# Tuning Learners:
library("mlr3tuning")
tuner <- tnr("intermbo")
# Tuning Objectives
library("bbotk")
optimizer <- opt("intermbo")
The tuner / optimizer provide an extensiveParamSet
to configure the MBO method, covering practically everything that can usually be configured with an MBOControl
object. To find out the specific function of each control parameter, read the mlrMBO reference entries of functions regarding "mlrMBO Control".
Installing
When installing mlrintermbo
, the required mlrMBO
package is not installed automatically. It is therefore necessary to install mlrMBO
manually:
install.packages("mlrMBO")
install.packages("mlrintermbo")
Known Issues
Assertion on 'xdt' failed
Error in .__OptimInstance__eval_batch(self = self, private = private, : Assertion on 'xdt' failed: Must have at least 1 rows, but has 0 rows.
This is caused by a bug in the
callr
package. The bug is fixed on CRAN, installing the current versions usinginstall.packages(c("callr", "processx"))
should fix the issue.Most other errors
Some errors, for example
Error: Domains[,1] must be less than or equal to Domains[,2]
are caused because the surrogate model is failing. (The error above happens when the objective function is giving constant values, which the surrogate learner does not handle well). Initialize the tuner with
on.surrogate.error
set to"warn"
or"quiet"
to ignore errors of the surrogate model. E.g.:tuner <- tnr("intermbo", on.surrogate.error = "warn") # alternatively "quiet"
License
LGPL-3