Published: 1997 Mai
Institution: University of Karlsruhe, Institute AIFB
Erscheinungsort / Ort: 76128 Karlsruhe, Germany
An empirical evaluation of the ILP system JoJo-FOL is presented. JoJo-FOL uses an iterative method to discover rules for a single concept within the normal inductive logic programming framework. The search process starts with a very restricted hypothesis space to efficiently derive most general and correct clauses first. This strong language bias is weakened by iteratively increasing the number of variables allowed to occur in a clause such that JoJo-FOL is able to induce additional correct but more specific clauses. JoJo-FOL conducts its search through the hypothesis space in a bidirectional manner controlled by two criteria of success derived from the completeness and consistency requirements. In an empirical evaluation JoJo-FOL was tested on two ILP benchmark data sets. The results presented show that the iterative rule discovery method combined with the bidirectional search strategy leads to more general rules and a higher accuracy compared to well-known ILP systems on the tested domains.