LoginSign Up
← Back to all sparks
M

mlr3tuningspaces

AI-ASSISTANTS
Velocity2.5

Collection of search spaces for hyperparameter optimization in the mlr3 ecosystem

github.com/mlr-org/mlr3tuningspaces ↗

A curated catalogue of published hyperparameter search spaces, now reaching deep neural networks

hyperparameter-tuningmlr3benchmark-studiesr-packagemachine-learning
◆Current state
mlr3tuningspaces packages hyperparameter search spaces taken from published benchmark studies so mlr3 users can tune against a citable range instead of inventing bounds. Its release history is steady catalogue growth punctuated by compatibility bumps across the mlr3 stack. 0.7.0 adds spaces for deep neural networks from Gorishniy, Rubachev, Khrulkov and Babenko (2021) alongside mlr3 1.7.2 compatibility.
◆Where it's heading
The catalogue keeps widening one paper at a time — Kühn (2018) rbv1 spaces in 0.4.0, a corrected attribution to Binder, Pfisterer and Bischl (2020) for rbv2 in the same release, and now a deep-learning set in 0.7.0. That growth is bounded by forces outside the package: 0.6.0 had to delete the `kknn` spaces outright when the underlying package left CRAN, a breaking change driven by upstream availability rather than any design decision here.
◆Prediction
Expect further spaces from newly published benchmark papers rather than a change in what the package does, since every feature release in this window has been of that form. Whether the deep-learning spaces get extended depends on learner support elsewhere in mlr3, which these entries do not cover.

◆Recent moves

  1. 21d ago

    Deep neural network tuning spaces added

    Tuning spaces for deep neural networks from Gorishniy et al. (2021) enter the catalogue, extending it past the classical learners it has covered so far, with mlr3 1.7.2 compatibility alongside. It is catalogue growth in the package's established pattern of adding one published study at a time.

    View source ↗
  2. 1y ago

    kknn tuning spaces removed after CRAN departure

    A breaking removal forced from outside: the `kknn` package left CRAN, so its tuning spaces are deleted from the catalogue. Printing also moves to the `cli` package. Anyone tuning k-nearest-neighbour learners through these spaces loses them with no in-package replacement.

    View source ↗
  3. 1y ago

    Compatibility with mlr3learners 0.9.0

    A compatibility-only release tracking mlr3learners 0.9.0, with no change to the catalogue itself.

    View source ↗
  4. 2y ago

    Compatibility with mlr3tuning 1.0.0

    Compatibility with mlr3tuning 1.0.0, part of the coordinated version bump across the mlr3 tuning stack.

    View source ↗
  5. 2y ago

    ranger.rbv1 factor handling narrowed; paradox 1.0.0 support

    Narrows `respect.unordered.factors` in the `ranger.rbv1` search space to `"ignore"` and `"order"`, plus paradox 1.0.0 compatibility. A correction inside an existing space rather than a new one.

    View source ↗
  6. 3y ago

    rbv1 search spaces added; rbv2 attribution corrected

    Adds the rbv1 search spaces from Kühn (2018) and corrects the attribution of the rbv2 spaces to Binder, Pfisterer and Bischl (2020). For a package whose value rests on citable provenance, fixing a wrong source is more than a documentation edit.

    View source ↗