Section: New Results

Force-Based Cooperative Search Directions in Evolutionary Multi-objective Optimization

Participants: Bilel Derbel, Dimo Brockhoff, Arnaud Liefooghe

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. In [46] , we 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 $\rho$MNK landscapes with different objective correlations. It turns out that the new approach is easy to implement, fast, and competitive with respect to a $(\mu +\lambda )$-SMS-EMOA variant, in particular if the objectives show strong positive or negative correlations.