Department of Computer Science Technische Universitaet Muenchen
Home People Teaching Research Publications Info Sitemap
Research | Publications | Teaching | Miscellaneous | Contact

Publications

Sorted by DateClassified by Publication TypeClassified by Research Category

Freek Stulp, Mark Pflüger, and Michael Beetz. Feature Space Generation using Equation Discovery. In Proceedings of the 29th German Conference on Artificial Intelligence (KI), 2006.

Download

[PDF]113.5kB  

Abstract

In Machine Learning, the success and performance of learning often critically depends on the feature space provided. This is particularly true for robot learning, where feature spaces are typically high dimensional and features are continuous. For many applications, it is common to design feature spaces manually. State variables are composed into higher level features using domain-specific knowledge. Unfortunately, manually designing these feature languages is tedious, because each new learning problem usually needs its own customized feature space. It is also error-prone, as relevant information in the original state space might be lost in the transformation. To overcome these problems, we propose an algorithm that automatically generates compact feature spaces, based on Equation Discovery. Our novel approach combines the strengths of Equation Discovery, being the compactness and interpretability of the resulting function, and Machine Learning, being its ability to approximate complex data.

BibTeX Entry

@InProceedings{stulp06feature,
  author =            {Freek Stulp and Mark Pfl\"uger and Michael Beetz},
  title =             {Feature Space Generation using Equation Discovery},
  booktitle =         {Proceedings of the 29th German Conference on Artificial Intelligence (KI)},
  year =              {2006},
  bib2html_pubtype =  {Refereed Conference Paper},
  bib2html_rescat =   {Robotics - Motor Control},
  abstract =          {In Machine Learning, the success and performance of learning often
  critically depends on the feature space provided. This is particularly true
  for robot learning, where feature spaces are typically high dimensional and
  features are continuous. For many applications, it is common to design feature
  spaces manually. State variables are composed into higher level features using
  domain-specific knowledge. Unfortunately, manually designing these feature
  languages is tedious, because each new learning problem usually needs its own
  customized feature space. It is also error-prone, as relevant information in
  the original state space might be lost in the transformation.
  To overcome these problems, we propose an algorithm that automatically
  generates compact feature spaces, based on Equation Discovery. Our novel
  approach combines the strengths of Equation Discovery, being the compactness
  and interpretability of the resulting function, and Machine Learning, being
  its ability to approximate complex data.}
}

Generated by bib2html.pl (written by Patrick Riley ) on Tue Sep 23, 2008 15:09:11 UTC


Computer Science IX | Technische Universität München