A Principled Evaluation of the Effect of Directed Mutation on Search-Based Statistical Testing

Statistical testing generates test inputs by sampling from a probability distribution that is carefully chosen so that the inputs exercise all parts of the software being tested. Sets of such inputs have been shown to detect more faults than test sets generated using traditional random and structural testing techniques. Search-based statistical testing employs a metaheuristic search algorithm to automate the otherwise labour-intensive process of deriving the probability distribution. This paper proposes an enhancement to this search algorithm: information obtained during fitness evaluation is used to direct the mutation operator to those parts of the representation where changes may be most beneficial. A principled empirical evaluation demonstrates that this enhancement leads to a significant improvement in algorithm performance, and so increases both the cost-effectiveness and scalability of search-based statistical testing. As part of the empirical approach, we demonstrate the use of response surface methodology as an effective and objective method of tuning algorithm parameters, and suggest innovative refinements to this methodology.