Correlation guided model building

Abstract

The intrinsic feature of Estimation of Distribution Algorithms lies in their ability to learn and employ probabilistic models over the input spaces. Discovery of the appropriate model usually implies a computationally expensive comprehensive search, where many models are proposed and evaluated in order to find the best value of some model discriminative scoring metric.This paper presents basic results demonstrating how simple variable correlation data can be extended and used to efficiently guide the model search, decreasing the number of model evaluations by several orders of magnitude and without significantly affecting model quality.As a case study, the O(n 3 ) model building of the Extended Compact Genetic Algorithm is successfully replaced by a correlation guided search of linear complexity which infers the perfect problem structures on the test suites.