Sie haben Javascript deaktiviert!
Sie haben versucht eine Funktion zu nutzen, die nur mit Javascript möglich ist. Um sämtliche Funktionalitäten unserer Internetseite zu nutzen, aktivieren Sie bitte Javascript in Ihrem Browser.

Info-Icon Diese Seite ist nicht in Deutsch verfügbar

CRC 901 – On-The-Fly Computing (OTF Computing)

ML-Plan - AutoML via Hierarchical Planning

ML-Plan is a state-of-the-art tool for automated machine learning (AutoML). In AutoML, the goal is to find the best possible combination of machine learning algorithms for a given learning problem while simultaneously optimizing their hyperparameters. In order to tailor the algorithms to the given learning problem, candidate solutions have to be executed and evaluated repeatedly during the optimization process.

ML-Plan tackles this problem by leveraging a technique called HTN and dividing the process into two phases. While the first phase explores the solution space of possible candidates with a reduced amount of data by means of a heuristic search, in the second phase a selection of the most promising observed solutions is re-evaluated with the addition of the retained data. Thus the return of a solution, which may be too specialized on the training data used due to an extensive search within the first phase, can be avoided.

In addition to the highly competitive and sometimes even superior generalizaton performance of solutions returned by ML-Plan compared to the ones of related tools, ML-Plan can be of advantage with respect to additional properties:

  • ML-Plan is currently the only AutoML tool that works with different machine learning libraries and can combine algorithms from different libraries using service-oriented architecture. In particular, the algorithms may be implemented in different programming languages.
  • Furthermore, ML-Plan is flexible and extensible, supports the addition and removal of available algorithms and the introduction of complex composition structures without having to make changes to ML-Plan itself.
  • ML-Plan scales with the available resources and enables a high degree of parallelization of the search by distributing the execution of candidates over several calculation nodes.



If you have any questions regarding ML-Plan, please, contact the staff of subproject B2.

Die Universität der Informationsgesellschaft