Investigating Problem Hardness in (Multi-Objective) Combinatorial Optimization: Algorithm Selection, Instance Generation and Tailored Algorithm Design

Optimization plays a crucial role in most fields of application and science. Thus, there is a need for sophisticated algorithms that tackle optimization problems and produce high quality solutions. In this context, research invented a plethora of algorithmic approaches. Evolutionary algorithms, i. e., heuristics, which mimick principles from Darwinian theory of evolution, are among the most active and rapidly growing fields of research and prove particularly effective in multi-objective optimization and, in general, for NP-hard combinatorial optimization problems. Often, many competing algorithms exist and it is a priori unclear which algorithm is the best choice for a given problem instance. The field of algorithm selection is a hot-topic with respect to this issue. It tries to build predictive models by means of computational intelligence, which utilize structural differences between instances to come up with appropriate algorithmic choices. My work in this cumulative thesis addresses issues from both pillars, i. e., evolutionary algorithm development and algorithm selection with focus on combinatorial optimization problems. I contribute to knowledge-discovery and expert-knowledge injection / biased variation operator design, local-search in multi-objective evolutionary algorithms and investigate various aspects of per-instance algorithm selection (application, performance measurement, discriminating feature identification). Statistical data analytics methods play a major role in most contributions to gain new knowledge and leverage the latter in order to solve problems. Last but not least accompanying software is presented that aided / enabled my research. Here, a special focus is put on a white-box framework for evolutionary algorithms.