Structure-Preserving Instance Generation.

Real-world instances are critical for the development of state-of-the-art algorithms, algorithm configuration techniques, and selection approaches. However, very few true industrial instances exist for most problems, which poses a problem both to algorithm designers and methods for algorithm selection. The lack of enough real data leads to an inability for algorithm designers to show the effectiveness of their techniques, and for algorithm selection it is difficult or even impossible to train a portfolio with so few training examples. This paper introduces a novel instance generator that creates instances that have the same structural properties as industrial instances. We generate instances through a large neighborhood search-like method that combines components of instances together to form new ones. We test our approach on the MaxSAT and SAT problems, and then demonstrate that portfolios trained on these generated instances perform just as well or even better than those trained on the real instances.