Search-based Unit Test Generation for Evolving Software

Search-based software testing has been successfully applied to generate unit test cases for object-oriented software. Typically, in search-based test generation approaches, evolutionary search algorithms are guided by code coverage criteria such as branch coverage to generate tests for individual coverage objectives. Although it has been shown that this approach can be effective, there remain fundamental open questions. In particular, which criteria should test generation use in order to produce the best test suites? Which evolutionary algorithms are more effective at generating test cases with high coverage? How to scale up search-based unit test generation to software projects consisting of large numbers of components, evolving and changing frequently over time? As a result, the applicability of search-based test generation techniques in practice is still fundamentally limited. In order to answer these fundamental questions, we investigate the following improvements to search-based testing. First, we propose the simultaneous optimisation of several coverage criteria at the same time using an evolutionary algorithm, rather than optimising for individual criteria. We then perform an empirical evaluation of different evolutionary algorithms to understand the influence of each one on the test optimisation problem. We then extend a coverage-based test generation with a non-functional criterion to increase the likelihood of detecting faults as well as helping developers to identify the locations of the faults. Finally, we propose several strategies and tools to efficiently apply search-based test generation techniques in large and evolving software projects. Our results show that, overall, the optimisation of several coverage criteria is efficient, there is indeed an evolutionary algorithm that clearly works better for test generation problem than others, the extended coverage-based test generation is effective at revealing and localising faults, and our proposed strategies, specifically designed to test entire software projects in a continuous way, improve efficiency and lead to higher code coverage. Consequently, the techniques and toolset presented in this thesis - which provides support to all contributions here described - brings search-based software testing one step closer to practical usage, by equipping software engineers with the state of the art in automated test generation