SCIENCE@home

Universal Learning of Repeated Matrix Games

J. Poland,Marcus Hutter

Published 2005 in arXiv.org

ABSTRACT

AbstractWe study and compare the learning dynamics of two universal learningalgorithms, one based on Bayesian learning and the other on prediction withexpert advice. Both approaches have strong asymptotic performance guaran-tees. When confronted with the task of finding good long-term strategies inrepeated 2 ×2 matrix games, they behave quite differently. 1 Introduction Today, Data Mining and Machine Learning is typically treated in a problem-specific way: People propose algorithms to solve a particular problem (such as learning toclassify points in a vector space), they prove properties and performance guaran-tees of their algorithms (e.g. for Support Vector Machines), and they evaluate thealgorithms on toy or real data, with the (potential) aim to use them afterwards inreal-world applications. In contrast, it seems that universal learning , i.e. a single al-gorithm which is applied for all (or at least “many”) problems, is neither feasible interms of computational costs nor competitive in (practical) performance. Neverthe-less, understanding universal learning is important: On the one hand, its practicalsuccess would lead a way to Artificial Intelligence. On the other hand,

PUBLICATION RECORD

  • Publication year

    2005

  • Venue

    arXiv.org

  • Publication date

    2005-08-16

  • Fields of study

    Mathematics, Computer Science

  • Identifiers
    arXiv cs/0508073
  • External record

    Open on Semantic Scholar

  • Source metadata

    Semantic Scholar

CITATION MAP

EXTRACTION MAP

CLAIMS

  • No claims are published for this paper.

CONCEPTS

  • No concepts are published for this paper.

REFERENCES

Showing 1-17 of 17 references · Page 1 of 1

CITED BY

Showing 1-18 of 18 citing papers · Page 1 of 1