Tabata: A Learning Algorithm Performing a Bidirectional Search in a Reduced Search Space Using a Tabu Strategy.
Pierre Brézellec, Henry Soldano · 1998
In concept learning from instances, a partially ordered concept space is searched. Whereas top-down methods search the whole space, bottom-up methods search a reduced space whose elements are most specific generalizations of positive instances. In this work we give the foundations for this reduction, and propose a method that performs a bidirectional search of this reduced space using a Tabu search strategy. 1 INTRODUCTION Here we consider concept learning from instances of disjoint classes. Most symbolic methods address this problem by searching, for each class, a " best " concept definition, according to some preference criteria, within a concept space. The instances of the target class are seen as the positive ones and the others as the negative ones. In rule learning methods, an element of the concept space is a set of conjunctive terms, each of which represents the left part of a rule. As the concept space (called E throughout this paper) is partially ordered, most methods perf...