Multi-Modal Optimisation using a Localised Surrogates Assisted Evolutionary Algorithm

Abstract

There has been a steady growth in interest in niching approaches within the evolutionary computation community, as an increasing number of real world problems confronted are discovered that exhibit multi-modality of varying degrees of intensity (modes). The presence of multi-modality can lead to some optimisers stalling on a sub-optimal mode – however it is often also useful to locate and memorise the modes encountered. This is because the optimal decision parameter combinations discovered may not actually be feasible when moving from a mathematical model emulating the real problem, to engineering an actual solution. As such a range of disparate modal solutions is of practical use. This paper investigates the use of a collection of localised surrogate models for niche/mode discovery, and analyses the performance of a novel evolutionary algorithm which embeds these surrogates into its search process. Results obtained are compared to published performance of state-of-the- art evolutionary algorithms developed for multi-modal problems. We find that using a collection of localised surrogates not only makes the problem tractable from a model-fitting viewpoint, it also produces competitive results with other EA approaches.