A Comparison of Dimensionality Reduction Techniques for Visualising Search Behaviour

When comparing different metaheuristics, the behaviour of the search process is a central feature towards distinguishing them beyond their mere metaphors. However, with large-scale studies, i.e. featuring hundreds or more configurations on large sets of benchmark functions, which are important to assess true capabilities, the sheer number of runs, high-dimensional data, and complex presentations make the traditionally employed visual evaluation of behaviours impossible. To remedy this, techniques that can summarise data without loosing important information are vital. In this paper, we investigate the effectiveness of five established techniques from the field of unsupervised learning in an example study featuring configurations of random search, genetic algorithms, and particle swarm optimisation on a set of optimisation benchmark functions. We find that the data reduction techniques do not perform equally well and that they can not be used to their full effect with default parameters. However, we demonstrate that these techniques can provide valuable insights into algorithmic behaviour.