Tutorial V many-objective evolutionary algorithms for optimization

Abstract

Evolutionary computation is the study of biologically motivated computational paradigms which exert novel ideas and inspiration from natural evolution and adaptation. The applications of population-based heuristics and nature-inspired metaphors in solving multiobjective optimization problems have been receiving a growing attention. To search for a family of Pareto optimal solutions, Evolutionary Multiobjective Optimization Algorithms have been successfully exploited to solve optimization problems in which the fitness measures and even constraints could be uncertain and varied over time. When encounter optimization problems with many objectives, nearly all designs performs poorly because of loss of selection pressure in fitness evaluation solely based upon Pareto optimality principle. This tutorial will survey recently published literature along this line of research- evolutionary algorithm for many-objective optimization and its real-world applications. Specifically, selection strategy, including mating selection and environmental selection, is a key ingredient in the design of evolutionary many-objective optimization algorithms. We will provide a comprehensive analysis on the selection strategies in the current evolutionary many-objective optimization algorithms. Experimental results on scalable DTLZ and WFG benchmark functions will demonstrate the pros and cons of various designs in terms of chosen performance metrics designed specifically for many-objective optimization. Based on performance metrics ensemble, we will provide a comprehensive measure among all competitors and more importantly reveal insight pertaining to specific problem characteristics that each evolutionary many-objective optimization algorithm could perform the best. The experimental results confirm the finding from the No Free Lunch theorem that any algorithm's elevated performance over one class of problems is exactly paid for in loss over another class.