Scholarlink

Test Case Prioritization (TCP) is crucial for efficient regression testing in Continuous Integration (CI). Existing reinforcement learning (RL)-based TCP methods face two limitations: (1) New test cases lack historical data for initial prioritization, leading to unstable early-stage sorting; (2) Over-reliance on short-term feedback limits cross-cycle behavior modeling. To address these, we propose LARA-TCP, integrating three key components: (1) An initial priority module that assigns non-random priorities to new test cases using intrinsic features (verdict, duration, error count); (2) A reward allocation module combining LSTM networks and attention mechanisms to model cross-cycle dependencies; (3) A multi-armed bandit (MAB) scheduler balancing exploration and exploitation. Experiments on Druid, IOF/ROL, and Retrofit datasets demonstrate LARA-TCP’s superiority under varying time budgets (10%-80%). Key findings: LARA-TCP achieves higher fault detection coverage (measured by NAPFD/APFDc) than baseline methods, particularly under constrained budgets (e.g., 10% time). Ablation studies confirm the individual contributions of the initial priority and reward allocation modules. The method shows consistent adaptability across datasets with diverse test suite characteristics. LARA-TCP provides a robust, history-aware prioritization strategy for CI environments, effectively addressing initialization bias and short-term feedback dependency.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

RL-Based Test Case Prioritization with LSTM-Attention Reward Function

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Zheng Su,
  • Minghong Luo,
  • Yan Tang,
  • Qiang Chen,
  • Xiaoming Ding

摘要

Test Case Prioritization (TCP) is crucial for efficient regression testing in Continuous Integration (CI). Existing reinforcement learning (RL)-based TCP methods face two limitations: (1) New test cases lack historical data for initial prioritization, leading to unstable early-stage sorting; (2) Over-reliance on short-term feedback limits cross-cycle behavior modeling. To address these, we propose LARA-TCP, integrating three key components: (1) An initial priority module that assigns non-random priorities to new test cases using intrinsic features (verdict, duration, error count); (2) A reward allocation module combining LSTM networks and attention mechanisms to model cross-cycle dependencies; (3) A multi-armed bandit (MAB) scheduler balancing exploration and exploitation. Experiments on Druid, IOF/ROL, and Retrofit datasets demonstrate LARA-TCP’s superiority under varying time budgets (10%-80%). Key findings: LARA-TCP achieves higher fault detection coverage (measured by NAPFD/APFDc) than baseline methods, particularly under constrained budgets (e.g., 10% time). Ablation studies confirm the individual contributions of the initial priority and reward allocation modules. The method shows consistent adaptability across datasets with diverse test suite characteristics. LARA-TCP provides a robust, history-aware prioritization strategy for CI environments, effectively addressing initialization bias and short-term feedback dependency.