Edinburgh Napier University

  In this chapter, we consider, formalise, and demonstrate the ways in which XAI can assist or inform algorithm selection and configuration. Reviewing the literature, we notice and taxonomise a much broader and more diverse notion of XAI than is typically considered for these purposes. Thereafter, the chapter includes two case studies which each demonstrate approaches from the taxonomy: in the first, a mixture of standard XAI and non-standard XAI methods are used to understand and explain algorithmic structural bias in a large set of CMA-ES configurations on the BBOB benchmark and to examine how the bias manifests on different landscapes. In the second, XAI techniques are applied to understand the limitations and robustness of an algorithm selection model in discrete optimisation. To this end, “adversarial” problem instances are evolved with the aim of causing well-performing algorithm selection models to mis-classify. The attributes of these adversarial instances are examined, bringing insight into the model and its training data. We conclude the chapter by outlining our suggestions for how methods from the taxonomy should be used in the future, highlighting under-researched avenues, and providing our outlook on XAI for algorithm selection and configuration.