Improving Readability in Automatic Unit Test Generation

In object-oriented programming, quality assurance is commonly provided through writing unit tests, to exercise the operations of each class. If unit tests are created and maintained manually, this can be a time-consuming and laborious task. For this reason, automatic methods are often used to generate tests that seek to cover all paths of the tested code. Search may be guided by criteria that are opaque to the programmer, resulting in test sequences that are long and confusing. This has a negative impact on test maintenance. Once tests have been created, the job is not done: programmers need to reason about the tests throughout the lifecycle, as the tested software units evolve. Maintenance includes diagnosing failing tests (whether due to a software fault or an invalid test) and preserving test oracles (ensuring that checked assertions are still relevant). Programmers also need to understand the tests created for code that they did not write themselves, in order to understand the intent of that code. If generated tests cannot be easily understood, then they will be extremely difficult to maintain. The overall objective of this thesis is to reaffirm the importance of unit test maintenance; and to offer novel techniques to improve the readability of automatically generated tests. The first contribution is an empirical survey of 225 developers from different parts of the world, who were asked to give their opinions about unit testing practices and problems. The survey responses confirm that unit testing is considered important; and that there is an appetite for higher-quality automated test generation, with a view to test maintenance. The second contribution is a domain-specific model of unit test readability, based on human judgements. The model is used to augment automated unit test generation to produce test suites with both high coverage and improved readability. In evaluations, 30 programmers preferred our improved tests and were able to answer maintenance questions 14level of accuracy. The third contribution is a novel algorithm for generating descriptive test names that summarise API- level coverage goals. Test optimisation ensures that each test is short, bears a clear relation to the covered code, and can be readily identified by programmers. In evaluations, 47 programmers agreed with the choice of synthesised names and that these were as descriptive as manually chosen names. Participants were also more accurate and faster at matching generated tests against the tested code, compared to matching with manually-chosen test names.