Meta-learning based selection of software reliability models

摘要

The correct estimation of the software reliability level is fundamental to reduce efforts in the testing, maintenance and release activities. To help in this task, we find in the literature an increasing number of software reliability models (SRMs). However, none has proved to perform well considering different contexts. Due to this, the selection of the best model for a particular case is an important task. Most existing works on SRM selection need to test different models and decide based on how well the model fits the data and predicts the future events. Moreover, in general, they do not consider search-based models. Considering this fact, this paper introduces a Meta-learning approach for SRM selection. In such approach, some meta-features are used to indicate the best performing model. The approach is independent of the type of models to be selected, and can be used with different data mining algorithms. It includes the following activities: meta-knowledge extraction, meta-learning and classification. The activities meta-knowledge extraction and meta-learning are performed just once and generate a meta-classifier. Therefore, the meta-classifier is used to select the most adequate model for new projects (classification activity). The approach is evaluated in a set of experiments and the results do not show statistical difference between the Meta-learning approach and the choice of the best performing model. Otherwise, the results point out statistical difference between the Meta-learning approach and the choice of the worst performing model with a large stochastic difference according to the Vargha and Delaney Effect Size.