Automated Machine Learning for 'mlr3'

Flexible automated machine learning (AutoML) system for the 'mlr3' ecosystem. Automatically selects a suitable machine learning algorithm and tunes its hyperparameters for a given task. Constructs preprocessing pipelines with multiple parallel branches using 'mlr3pipelines' and jointly optimizes them together with the learners using 'mlr3tuning'. The optimization is driven by asynchronous decentralized Bayesian optimization by Egele et al. (2023) .

mlr3automl

Package website: release | dev

mlr3automl is the Automated Machine Learning (AutoML) package of the mlr3 ecosystem. It automatically selects the most suitable machine learning algorithm and tunes its hyperparameters for a given task. The package includes 10 learners from the mlr3learners package, ranging from simple models like glmnet to more powerful algorithms such as ranger and xgboost. Leveraging the mlr3pipelines package, it constructs sophisticated preprocessing graphs with multiple parallel branches, which are jointly optimized using the mlr3tuning package. The optimization is driven by Asynchronous Decentralized Bayesian Optimization (ADBO), enabling efficient and scalable AutoML.

Installation

Install the last release from CRAN:

install.packages("mlr3automl")

Install the development version from GitHub:

pak::pkg_install("mlr-org/mlr3automl")

Examples

library("mlr3automl")

rush_plan(n_workers = 2, worker_type = "mirai")
mirai::daemons(2)

task = tsk("spam")

learner = lrn("classif.auto",
  terminator = trm("evals", n_evals = 100)
)

learner$train(task)

Test with Redis

To test the package, set the RUSH_TEST_USE_REDIS environment variable to true. The test suite deletes the Redis database before execution, so never run it against a production server.

Sys.setenv(RUSH_TEST_USE_REDIS = "true")