Stochastically Guided Disjunctive Version Space Learning.
N.I. Nikolaev, Evgueni N. Smirnov · 1996
This paper presents an incremental concept learning approach to identification of concepts with high overall accuracy. The main idea is to address the concept overlap as a central problem when learning multiple descriptions. Many traditional inductive algorithms, as those from the disjunctive version space family considered here, face this problem. The approach focuses on combinations of confident, possibly overlapping, concepts with an original stochastic complexity formula. The focusing is efficient because it is organized as a simulated annealing-based beam search. The experiments show that the approach is especially suitable for developing incremental learning algorithms with the following advantages: first, it generates highly accurate concepts; second, it overcomes to a certain degree the sensitivity to the order of examples; and third, it handles noisy examples.