Multi-objective LTGA

Abstract

In recent years new types of evolutionary algorithms have been introduced that try to improve traditional crossover operators by learning the structure of the problem, most notably the Linkage Tree Genetic Algorithm. For single-objective problems linkage learning has proven to be a valuable improvement over the traditional crossover operators, since it can automatically learn and exploit the structure of a specific problem. Some initial research has been done to extend the use of LTGA into the domain of multi-objective problems, but no research has been done that looks at the effectiveness of linkage learning in a multi-objective environment. In order to do this some new multi-objective variants of the NK-landscape are introduced. For single-objective problems there are plenty of benchmark functions that model different types of problem variable correlations, but adaptations of these benchmark functions for multi-objective environments often leave the variable correlations between objectives unchanged.