Evaluate Software Quality by Learning from Historical Data

This paper proposes a method to evaluate software quality by learning from historical data. Quantitative evaluation of software quality is not an easy issue. But historical software can provide us with a lot of software quality information. We present a data acquiring model to guide the data collection from historical software. Then machine learning algorithm that support incremental training is applied to learn the relationship between software quality and software metrics from the data. As a case study, we collected the data of 82 aviation embedded software in an institute, and trained a k-Nearest Neighbors (k-NN) classification model optimized by genetic algorithms for evaluating the software reliability.