Applications of Model Reuse When Using Estimation of Distribution Algorithms to Test Concurrent Software

Abstract

Previous work has shown the efficacy of using Estimation of Distribution Algorithms (EDAs) to detect faults in concurrent software/systems. A promising feature of EDAs is the ability to analyse the information or model learned from any particular execution. The analysis performed can yield insights into the target problem allowing practitioners to adjust parameters of the algorithm or indeed the algorithm itself. This can lead to a saving in the effort required to perform future executions, which is particularly important when targeting expensive fitness functions such as searching concurrent software state spaces. In this work, we describe practical scenarios related to detecting concurrent faults in which reusing information discovered in EDA runs can save effort in future runs, and prove the potential of such reuse using an example scenario. The example scenario consists of examining problem families, and we provide empirical evidence showing real effort saving properties for three such families.