Scholarlink

In this paper, we present a case study of two safety-critical automotive software systems. We demonstrate how applying a data balancing technique to the training dataset significantly enhanced a machine learning-based verification strategy predictor, enabling it to formally verify 55% more property specifications across the two systems. A verification strategy is a sequence of techniques ordered in the decreasing likelihood of their ability to verify a given program with respect to a property specification. Such a strategy is needed because no single technique can verify all programs, and the availability of limited resources mandates an early prediction of suitable techniques in the strategy. To the best of our knowledge, this is the first study to explore the role of data balancing in the context of verification strategy prediction.

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

Random Resampling of Training Data for Effective Verification Strategy Prediction

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Bharti Chimdyalwar,
  • Priyanka Darke,
  • R. Venkatesh,
  • Supratik Chakraborty

摘要

In this paper, we present a case study of two safety-critical automotive software systems. We demonstrate how applying a data balancing technique to the training dataset significantly enhanced a machine learning-based verification strategy predictor, enabling it to formally verify 55% more property specifications across the two systems. A verification strategy is a sequence of techniques ordered in the decreasing likelihood of their ability to verify a given program with respect to a property specification. Such a strategy is needed because no single technique can verify all programs, and the availability of limited resources mandates an early prediction of suitable techniques in the strategy. To the best of our knowledge, this is the first study to explore the role of data balancing in the context of verification strategy prediction.