Advanced statistical analysis of empirical performance scaling

Theoretical running time complexity analysis is a widely adopted method for studying the scaling behaviour of algorithms. However, theoretical analysis remains intractable for many high-performance, heuristic algorithms. Recent advances in statistical methods for empirical running time scaling analysis have shown that many state-of-the-art algorithms can achieve significantly better scaling in practice than expected. However, current techniques have only been successfully applied to study algorithms on randomly generated instance sets, since they require instances that can be grouped into "bins", where each instance in a bin has the same size. In practice, real-world instance sets with this property are rarely available. We introduce a novel method that overcomes this limitation. We apply our method to a broad range of scenarios and demonstrate its effectiveness by revealing new insights into the scaling of several prominent algorithms; e.g., the SAT solver lingeling often appears to achieve sub-polynomial scaling on prominent bounded model checking instances, and the training times of scikit-learn's implementation of SVMs scale as a lower-degree polynomial than expected (≈ 1.51 instead of 2).