Learning as Optimization: Stochastic Generation of Multiple Knowledge

Abstract

Learning of rules can be stated as an optimization problem. While current learning algorithms use variants of hill climbing and beam search for combinatorial optimization stochastic algorithms, such as simulated annealing, are used. In the paper five stochastic learning algorithms that differ in the amount of search are defined and applied to parity problems of various degrees, the problem of legal position in KRK chess endgame and two medical diagnostic problems. The results indicate, that stochastic search may provide a useful upgrade of current learning algorithms.