A surrogate-ensemble assisted expensive many-objective optimization

Abstract

The surrogate ensemble has been shown to have a good performance for assisting the evolutionary algorithms to solve computationally expensive single-objective problems. However, it has not been paid much attention to assist in multi-/many-objective optimization. In this paper, we propose to train multiple surrogate models to assist many-objective optimization algorithm for solving expensive many-objective problems. In the proposed method, some non-dominated solutions, which have been evaluated using the expensive objectives, will be utilized together with the current population to generate the offspring in order to prevent the population from moving forward to a bad region. In the model management, the way to determine the method used for selecting individuals to be evaluated using the exact expensive objectives depends on the convergence degree of the current population. A new method to measure the uncertainty is also proposed in our method by considering the distance to the samples in the decision space and the approximation variance in the objective space. The experimental results on a number of benchmark problems show that our proposed method is competitive to two state-of-the-art surrogate-assisted evolutionary algorithms for solving expensive many-objective problems in a limited computational budget.