Ranking-prediction based evolutionary algorithm for expensive many-objective optimization problems
Abstract
To deal with challenging expensive many-objective optimization problems, this study proposes a novel ranking-prediction based evolutionary algorithm. Instead of approximating the objectives directly, our algorithm introduces M symmetrical ranking cases in the generalized Pareto dominance to construct a single comprehensive Kriging model. Moreover, we design a novel method referred to as the expected ranking advantage (ERA) for model management to cooperate with ranking-prediction, which is capable of providing sufficient information and avoiding the accumulation of errors compared to traditional surrogate models. In addition, our proposed algorithm achieves a significant improvement in the time complexity of model construction. The outstanding performance and efficiency of the ERA-based algorithm are demonstrated in two benchmark test suites with the comparison of four state-of-the-art algorithms.
Authors: Yimo Zhang, Shuwei Zhu, Wei Hua Fang, Kalyanmoy Deb, Meiji Cui
Published in: Genetic and Evolutionary Computation Conference Companion (GECCO Companion) (2025)