TY - GEN
T1 - Force-based cooperative search directions in evolutionary multi-objective optimization
AU - Derbel, Bilel
AU - Brockhoff, Dimo
AU - Liefooghe, Arnaud
PY - 2013/4/3
Y1 - 2013/4/3
N2 - In order to approximate the set of Pareto optimal solutions, several evolutionary multi-objective optimization (EMO) algorithms transfer the multi-objective problem into several independent single-objective ones by means of scalarizing functions. The choice of the scalarizing functions' underlying search directions, however, is typically problem-dependent and therefore difficult if no information about the problem characteristics are known before the search process. The goal of this paper is to present new ideas of how these search directions can be computed adaptively during the search process in a cooperative manner. Based on the idea of Newton's law of universal gravitation, solutions attract and repel each other in the objective space. Several force-based EMO algorithms are proposed and compared experimentally on general bi-objective ρMNK landscapes with different objective correlations. It turns out that the new approach is easy to implement, fast, and competitive with respect to a (μ + λ)-SMS-EMOA variant, in particular if the objectives show strong positive or negative correlations.
AB - In order to approximate the set of Pareto optimal solutions, several evolutionary multi-objective optimization (EMO) algorithms transfer the multi-objective problem into several independent single-objective ones by means of scalarizing functions. The choice of the scalarizing functions' underlying search directions, however, is typically problem-dependent and therefore difficult if no information about the problem characteristics are known before the search process. The goal of this paper is to present new ideas of how these search directions can be computed adaptively during the search process in a cooperative manner. Based on the idea of Newton's law of universal gravitation, solutions attract and repel each other in the objective space. Several force-based EMO algorithms are proposed and compared experimentally on general bi-objective ρMNK landscapes with different objective correlations. It turns out that the new approach is easy to implement, fast, and competitive with respect to a (μ + λ)-SMS-EMOA variant, in particular if the objectives show strong positive or negative correlations.
U2 - 10.1007/978-3-642-37140-0_30
DO - 10.1007/978-3-642-37140-0_30
M3 - Conference contribution
AN - SCOPUS:84875490879
SN - 9783642371394
T3 - Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
SP - 383
EP - 397
BT - Evolutionary Multi-Criterion Optimization - 7th International Conference, EMO 2013, Proceedings
T2 - 7th International Conference on Evolutionary Multi-Criterion Optimization, EMO 2013
Y2 - 19 March 2013 through 22 March 2013
ER -