Using an adaptive collection of local evolutionary algorithms for multi-modal problems

Abstract

Multi-modality can cause serious problems for many optimisers, often resulting convergence to sub-optimal modes. Even when this is not the case, it is often useful to locate and memorise a range of modes in the design space. This is because “optimal" decision parameter combinations may not actually be feasible when moving from a mathematical model emulating the real problem, to engineering an actual solution, making a range of disparate modal solutions of practical use. This paper builds upon our work on the use of a collection of localised search algorithms for niche/mode discovery which we presented at UKCI 2013 when using a collection of surrogate models to guide mode search. Here we present the results of using a collection of exploitative local evolutionary algorithms (EAs) within the same general framework. The algorithm dynamically adjusts its population size according to the number of regions it encounters that it believes contain a mode, and uses localised EAs to guide the mode exploitation. We find that using a collection of localised EAs, which have limited communication with each other, produces competitive results with the current state-of-the-art multimodal optimisation approaches on the CEC 2013 benchmark functions.