An Empirical Comparison of Several Recent Multi-objective Evolutionary Algorithms

Many real-world problems can be formulated as multi-objective optimisation problems, in which many potentially conflicting objectives need to be optimized simultaneously. Multi-objective optimisation algorithms based on Evolutionary Algorithms (EAs) such as Genetic Algorithms (GAs) have been proven to be superior to other traditional algorithms such as goal programming. In the past years, several novel Multi-Objective Evolutionary Algorithms (MOEAs) have been proposed. Rather than based on traditional GAs, these algorithms extended other EAs including novel EAs such as Scatter Search and Particle Swarm Optimiser to handle multi-objective problems. However, to the best of our knowledge, there is no fair and systematic comparison of these novel MOEAs. This paper, for the first time, presents the results of an exhaustive performance comparison of an assortment of 5 new and popular algorithms on the DTLZ benchmark functions using a set of well-known performance measures. We also propose a novel performance measure called unique hypervolume, which measures the volume of objective space dominated only by one or more solutions, with respect to a set of solutions. Based on our results, we obtain some important observations on how to choose an appropriate MOA according to the preferences of the user.