LoginSign Up
← Back to all sparks
M

mlr3mbo

ANALYTICS
Velocity2.5

Model-based and Bayesian optimization toolbox for the mlr3 ecosystem

github.com/mlr-org/mlr3mbo ↗

mlr3mbo picked its defaults from a benchmark study, not from taste

bayesian-optimizationmlr3hyperparameter-tuningr-statsdefaultsbenchmarking
◆Current state
mlr3mbo does model-based and Bayesian optimisation for mlr3. Its 1.0.0 release added a dictionary of pre-built acquisition-function optimisers and, more consequentially, replaced the default surrogate, acquisition function and optimiser settings with values derived from a large-scale benchmark study. The releases since are corrections to the acquisition-optimiser path exposed by that new default configuration.
◆Where it's heading
The package has moved from a toolkit that expected users to assemble a Bayesian optimisation loop into one with a defensible default loop, and the recent fixes — warm-start sizing on multi-objective archives, silently discarded terminators, stale x_domain values — are the consequences of more people running the default path.
◆Prediction
Expect continued hardening of the acquisition-optimiser classes rather than new acquisition functions.

◆Recent moves

  1. 23d ago

    Acquisition optimiser fixes for warm starts and archives

    Stops acqo() silently discarding terminators and callbacks, makes warm-start sizing behave on multi-objective archives, and removes stale x_domain values from proposed candidates. Several of these changed results quietly before being fixed.

    View source ↗
  2. 3mo ago

    Dictionary lookup and restart-limit fixes

    Four narrow fixes around the acquisition-optimiser dictionary, restart limits and output transformations. Follow-up to the 1.0.0 defaults work.

    View source ↗
  3. 4mo ago

    rush 1.0.0 compatibility and Surrogate$check()

    A compatibility release for the rush backend with one small validation method added to Surrogate.

    View source ↗
  4. 5mo ago

    mlr3mbo 1.0.0 ships benchmark-derived default settings

    ⚡ SPARK

    The release that made the package usable without assembling the optimisation loop yourself. Everything since is fallout from users actually running those defaults.

    View source ↗
  5. 10mo ago

    mlr3learners 0.13.0 compatibility

    A single-line compatibility release.

    View source ↗
  6. 11mo ago

    Maintainer change and mlr3pipelines 0.9.0 upkeep

    Records a maintainer handover alongside test fixes and one input-transformation fix for mixed spaces. Immediately precedes the 1.0.0 push.

    View source ↗