No abstract

Software systems tend to become more and more configurable to satisfy the demands of their increasingly varied customers. Exhaustively testing the correctness of highly configurable software is infeasible in most cases because the space of possible configurations is typically colossal. This paper proposes addressing this challenge by (i) working with a representative sample of the configurations, i.e., a "uniform" random sample, and (ii) processing the results of testing the sample with a rule induction system that extracts the faults that cause the tests to fail. The paper (i) gives a concrete implementation of the approach, (ii) compares the performance of the rule learning algorithms AQ, CN2, LEM2, PART, and RIPPER, and (iii) provides empirical evidence supporting our procedure.