A Generative Adversarial Imitation Learning Method for Continuous Integration Testing

In Continuous Integration (CI), Test Case Prioritization (TCP) is crucial for the efficiency and effectiveness of the software testing. While Reinforcement Learning (RL) offers a promising approach for TCP, it struggles with the low failure rates of test cases in industrial CI environment, leading to sparse rewards and unstable learning efficiency. Furthermore, designing a proper reward function is challenging due to its dependency on the abstracted features of the test cases. To address these issues, we propose a Generative Adversarial Imitation Learning (GAIL) method for TCP, which allows agents to learn directly from the expert experience rather than through potentially biased reward functions. We use the Copeland method as a pairwise ranking strategy and train the agent using optimal rankings, considered as expert experience generated from previous CI cycles, leading to more stable and efficient learning. In addition, we introduce a new metric, the Average of the Percentage of Faults Detected based on Execution Time (APFDET), to evaluate the effectiveness of the proposed approach. Empirical studies are performed on six industrial datasets. The results show that GAIL has a better fault detection capability on TCP problems.