Surrogate-assisted analysis of the parameter configuration landscape for meta-heuristic optimisation

Abstract

Meta-heuristics can provide high-quality solutions to challenging problems in a reasonable amount of time, but are highly sensitive to the values assigned to their control parameters. The parameter configuration landscape (PCL) offers insight into the characteristics associated with optimisation of the parameter configuration of a meta-heuristic, but is poorly understood for most meta-heuristics. Further exacerbating this issue, determining the characteristics of the PCL is an extremely computationally expensive process. This study proposes the usage of artificial neural networks (ANNs) as surrogate models to greatly reduce the computational burden associated with characterising the PCL. Notably, this study represents the first usage of surrogate models in PCL research. Furthermore, this study presents a characterisation of the PCLs for both particle swarm optimization (PSO) and differential evolution (DE) employing ANN surrogate models, using five well-established fitness landscape analysis (FLA) metrics, and finds that the common assumption of correlation between the fitness and distance of control parameter settings is not strictly met. Overall, the training and usage of the surrogate models leads to a 99.86% reduction in the number of algorithm executions required to attain the PCL samples used in the characterisation.