previous  home  search  LaTeX  -  PostScript  -  PDF  -  Html/Gif   contact    up    next  

On the Convergence Speed of MDL Predictions for Bernoulli Sequences


Authors: Jan Poland and Marcus Hutter (2004)
Comments: 16 pages
Subj-class: Probability Theory; Learning; Artificial Intelligence; Information Theory; Complexity
Reference: Proceedings of the 15th International Conference on Algorithmic Learning Theory (ALT 2004) pages 294-308
Report-no: IDSIA-13-04 and cs.LG/0407039
Paper: LaTeX  -  PostScript  -  PDF  -  Html/Gif 
Slides: PowerPoint (needs tex4ppt) - PDF

Keywords: MDL, Minimum Description Length, Convergence Rate, Prediction, Bernoulli, Discrete Model Class.

Abstract: We consider the Minimum Description Length principle for online sequence prediction. If the underlying model class is discrete, then the total expected square loss is a particularly interesting performance measure: (a) this quantity is bounded, implying convergence with probability one, and (b) it additionally specifies a `rate of convergence'. Generally, for MDL only exponential loss bounds hold, as opposed to the linear bounds for a Bayes mixture. We show that this is even the case if the model class contains only Bernoulli distributions. We derive a new upper bound on the prediction error for countable Bernoulli classes. This implies a small bound (comparable to the one for Bayes mixtures) for certain important model classes. The results apply to many Machine Learning tasks including classification and hypothesis testing. We provide arguments that our theorems generalize to countable classes of i.i.d. models.

 previous  home  search  LaTeX  -  PostScript  -  PDF  -  Html/Gif   contact    up    next  

BibTeX Entry

@InProceedings{Hutter:04mdlspeed,
  author =       "J. Poland and M. Hutter",
  title =        "On the convergence speed of {MDL} predictions for {B}ernoulli sequences",
  booktitle =    "Proc. 15th International Conf. on Algorithmic Learning Theory ({ALT-2004})",
  address =      "Padova",
  series =       "LNAI",
  volume =       "3244",
  editor =       "S. Ben-David and J. Case and A. Maruoka",
  publisher =    "Springer, Berlin",
  pages =        "294--308",
  year =         "2004",
  url =          "http://www.hutter1.net/ai/mdlspeed.htm",
  http =         "http://arxiv.org/abs/cs.LG/0407039",
  ftp =          "ftp://ftp.idsia.ch/pub/techrep/IDSIA-13-04.pdf",
  keywords =     "MDL, Minimum Description Length, Convergence Rate,
                  Prediction, Bernoulli, Discrete Model Class.",
  abstract =     "We consider the Minimum Description Length principle for online
                  sequence prediction. If the underlying model class is discrete,
                  then the total expected square loss is a particularly interesting
                  performance measure: (a) this quantity is bounded, implying
                  convergence with probability one, and (b) it additionally
                  specifies a `rate of convergence'. Generally, for MDL only
                  exponential loss bounds hold, as opposed to the linear bounds for
                  a Bayes mixture. We show that this is even the case if the model
                  class contains only Bernoulli distributions. We derive a new upper
                  bound on the prediction error for countable Bernoulli classes.
                  This implies a small bound (comparable to the one for Bayes
                  mixtures) for certain important model classes. The results apply
                  to many Machine Learning tasks including classification and
                  hypothesis testing. We provide arguments that our theorems
                  generalize to countable classes of i.i.d. models.",
}
 previous  home  search  LaTeX  -  PostScript  -  PDF  -  Html/Gif   contact    up    next