Lightweight Interpolation-Based Surrogate Modelling for Multi-objective Continuous Optimisation

Abstract

We propose two surrogate-based strategies for increasing the convergence speed of multi-objective evolutionary algorithms (MOEAs) by stimulating the creation of high-quality individuals early in the run. Both offspring generation strategies are designed to leverage the fitness approximation capabilities of light-weight interpolation-based models constructed using an inverse distance weighting function. Our results indicate that for the two solvers we tested with, NSGA-II and DECMO2++, the application of the proposed strategies delivers a substantial improvement of early convergence speed across a test set consisting of 31 well-known benchmark problems.