Optimizing Test Case Prioritization With Meta Deep Reinforcement Learning in Continuous Integration

Abstract

Software developers use Continuous Integration (CI) environments to reduce integration issues and expedite development cycles. Regression testing is an important part of CI practice, as it includes reconducting all test cases to guarantee system stability after upgrades. However, as test suites grow, this process becomes increasingly resource-intensive and time-consuming. While many Test Case Prioritization (TCP) techniques have been proposed to address this challenge, previous approaches often rely on static configurations and lack the adaptability needed to handle the dynamic nature of CI environments and different dataset complexities. To address these gaps, this study presents a novel TCP framework based on Deep Reinforcement Learning (DRL), integrating a pairwise ranking model with state-of-the-art DRL algorithms, including A2C, DQN, PPO, and TRPO. The proposed framework improves prioritization accuracy and execution efficiency, particularly when combined with an optimal cycle count strategy. An adaptive training framework based on Meta-Deep Reinforcement Learning (meta-DRL) was introduced to further enhance the adaptability of the framework. This component allows the DRL agent to assess its performance during training and dynamically modify the essential hyperparameters, thereby enhancing its ability to develop successful prioritizing methods over time. Finally, the results of the proposed methodology demonstrate that Meta-Deep Reinforcement Learning (meta-DRL) significantly reduces the training time and achieves a 60% reduction compared to existing approaches. These findings show the efficiency of meta-DRL-based TCP in providing a scalable and adaptive solution for enhancing regression testing in CI environments.