UNI-MB - logo
UMNIK - logo
 
(UM)
  • Use of heuristics in empirical inductive logic programming
    Lavrač, Nada ; Cestnik, Bojan ; Džeroski, Sašo
    Inductive Logic Programming (ILP) has only recently addressed the problem of learning from noisy data. The main goal of the paper is to improve the understanding of noise-handling mechanisms by ... giving an analysis of proposed heuristics for dealing with noise. It is argued that the proposed accuracy and information gain heuristics can be used as search heuristics and stopping criteria in clause construction, as well as simplification criteria in post-procesing of clauses. Furthermore, these heuristics can be improved by applying latest advances in estimating probabilities from the distribution of covered positive and negative examples, in particular by using the m-estimate. The problem of learning illegal positions in a chess endgame is used to illustrate and analyse how different search heuristics split the training set according to the distribution of positive and negative examples, indicating that the m-estimate is most appropriate for predicting the best split.
    Vir: Proceedings ([17] str.)
    Vrsta gradiva - prispevek na konferenci
    Leto - 1992
    Jezik - angleški
    COBISS.SI-ID - 2723094

vir: Proceedings ([17] str.)

loading ...
loading ...
loading ...

Vnos na polico