Measuring and Maintaining Population Diversity in Search-Based Unit Test Generation

Genetic algorithms (GAs) have been demonstrated to be effective at generating unit tests. However, GAs often suffer from a loss of population diversity, which causes the search to prematurely converge, thus negatively affecting the resulting code coverage. One way to prevent premature convergence is to maintain and increase population diversity. Although the impact of population diversity on the performance of GAs is well-studied in the literature, little attention has been given to population diversity in unit test generation. We study how maintaining population diversity influences the Many-Objective Sorting Algorithm (MOSA), a state-of-the-art evolutionary search algorithm for generating unit tests. We define three diversity measures based on fitness entropy, test executions (phenotypic diversity), and Java statements (genotypic diversity). To improve diversity, we apply common methods that fall into two groups: niching (such as fitness sharing and clearing) and non-niching (such as diverse initial populations). Our results suggest that increasing diversity does not have a beneficial effect on coverage in general, but it may improve coverage once the search stagnates.