MNK-Landscapes における交叉の混合特性とアルゴリズム性能との関係

抄録

Landscape analysis and automated algorithm selection and tuning are still intensely studied topics in multi-objective optimisation. Landscape analysis captures a problem’s structure through numerical features, thereby supply-ing machine-learning models with rich information for making sound configuration decisions. Beyond this problem-centred view, it is also crucial to quantify—independently of the problem being solved—the intrinsic properties of an algorithm’s own components and to analyse how those properties affect performance. One such component is crossover. Widely investigated, the effectiveness of crossover depends on the operator’s ability to mix information, the specific characteristics of the problem, and the population diversity that emerges from the algorithm’s dynam-ics. This work concentrates on binary encodings and introduces a method for exploring the link between crossover features and the performance of a multi-objective evolutionary algorithm on problem subclasses whose variable in-teractions follow either random or nearest-neighbour patterns and whose scales differ. Using regression models, we identify the crossover features that are relevant to performance within each subclass and show how the influential features change as problem size varies.