An Improved Meta-heuristic Search for Constrained Interaction Testing

Combinatorial interaction testing (CIT) is a cost-effective sampling technique for discovering interaction faults in highly configurable systems. Recent work with greedy CIT algorithms efficiently supports constraints on the features that can coexist in a configuration. But when testing a single system configuration is expensive, greedy techniques perform worse than meta-heuristic algorithms because they produce larger samples. Unfortunately, current meta-heuristic algorithms are inefficient when constraints are present. We investigate the sources of inefficiency, focusing on simulated annealing, a well-studied meta-heuristic algorithm. From our findings we propose changes to improve performance, including a reorganized search space based on the CIT problem structure. Our empirical evaluation demonstrates that the optimizations reduce run-time by three orders of magnitude and yield smaller samples. Moreover, on real problems the new version compares favorably with greedy algorithms.