Improving test suite generation quality through machine learning-driven boundary value analysis

Abstract

Boundary value analysis (BVA) is a widely used method in software testing to identify errors at the boundaries of input domains. However, traditional BVA is resource intensive and often impractical for complex systems with expansive input spaces. Recent advances in machine learning offer potential for automating BVA, improving efficiency and fault detection capabilities. This paper introduces a machine learning based approach for the automatic generation of boundary test inputs. The research focuses on addressing the automation of BVA processes, with a particular emphasis on white-box testing scenarios. The proposed methodology consists of two main steps. First, a ML-based discriminator is trained to identify the existence of a boundary between two test inputs. Based on the discriminator’s output, we calculate the boundary density using two proposed methods: “pointDensity” and “pairDensity.” In the second step, Markov Chain Monte Carlo (MCMC) techniques are applied to generate test inputs guided by the calculated boundary densities. Experiments were conducted to evaluate the fault detection capabilities of the ML-based approach compared to concolic testing and manual boundary analysis. The results show that our proposed method surpasses manual boundary analysis in five of the eight programs and outperforms concolic testing in four out of eight programs.