Adaptive genetic Pareto ranking based on clustering

The proposed Pareto ranking scheme is meant for the selection of parents and survivors in multi-objective evolutionary optimizations. Commonly, the Pareto methods use just the dominance analysis in order to provide the partial sorting of solutions, without taking into account the specific strength of the conflict detected between objectives. This can generate undesired effects, such as the loss of diversity or the excessive spread of solutions induced by too weakly or too strongly conflicting criteria, respectively. For counteracting these disadvantages, the suggested approach adapts the ranking policies with respect to the distribution of the population in the objective space. The first innovation of the paper resides in the way in which the layout of the available solutions is examined. The analysis is based on clustering, followed by the Pareto-ranking of the resulted centers. The centers belonging to the best fronts are then used to depict the preferred searching area and to decide if the diversity of the solutions requires improvement. In this regard, the second contribution supports the diversification of the preferred solutions via rank adjustments. The suggested ranking algorithm is experimentally verified on several synthetic multi-objective optimizations and a multi-objective robot path planning. The testing scenarios exemplify different layouts of the Pareto fronts for diverse conflictive relationships between the two objectives.