Automatic Design of Hybrid Metaheuristics from Algorithmic Components

Metaheuristic algorithms are traditionally designed following a manual, iterative algorithm development process. While this process sometimes leads to high performing algorithms, it is labor-intensive, error-prone, difficult to reproduce, and explores only a limited number of design alternatives. In this article, we advocate the automatic design of hybrid metaheuristic algorithms. For an effective design process, we propose as main ingredients a unified view on metaheuristic algorithms, an effective implementation of the metaheuristic components but also of problem-specific algorithm components inside a flexible algorithm framework, and the exploitation of automated algorithm configuration techniques. With these ingredients we show that, for various, rather different combinatorial optimization problems, we can automatically generate high-performance metaheuristic algorithms that reach and in various cases surpass the performance of current state-of-the-art algorithms for the respective problems. Our results also indicate that a paradigm shift in how effective metaheuristic algorithms are designed is possible, which has significant advantages such as reproducibility, unification of existing approaches, and reduction in the development time of high-performing algorithms.