Multi-objective genetic algorithm with clustering-based ranking and direct control of diversity

This paper presents a new Pareto-ranking algorithm which can be used for solving multi-objective optimization problems with few objectives. The ranks are assigned by progressively combining the search with decision. More precisely, the decision is implemented via an adaptive clustering which guides the search towards the middle of the Pareto-front. This enables a gradual rejection of the solutions expected to be less useful for the application. By monitoring the evolutions of the depicted clusters, the procedure is able to detect premature convergence and to intervene for encouraging the preservation of population diversity. In this attempt, whenever necessary, a supplementary objective is added. It is meant to directly control the variety of the most valuable genetic material. The applicability of the approach is demonstrated on robot path planning, considering continuous working scenes with known non-convex and/or disjoint obstacles.