Abstract :
Recently, many automatic test generation techniques have been proposed, such as Randoop, Pex and jCUTE. However, usually test coverage of these techniques has been around 50-60% only, due to several challenges, such as 1) the object mutation problem, where test generators cannot create and/or modify test inputs to desired object states; and 2) the constraint solving problem, where test generators fail to solve path conditions to cover certain branches. By analyzing branches not covered by state-of-the-art techniques, we noticed that these challenges might not be so difficult for humans. To verify this hypothesis, we propose a Puzzle-based Automatic Testing environment (PAT) which decomposes object mutation and complex constraint solving problems into small puzzles for humans to solve. We generated PAT puzzles for two open source projects and asked different groups of people to solve these puzzles. It was shown that they could be effectively solved by humans: 231 out of 400 puzzles were solved by humans at an average speed of one minute per puzzle. The 231 puzzle solutions helped cover 534 and 308 additional branches (7.0% and 5.8% coverage improvement) in the two open source projects, on top of the saturated branch coverages achieved by the two state-of-the-art test generation techniques.
