Out-of-the-Box Prediction of Non-Functional Variant Properties Using Automated Machine Learning

A configurable system is characterized by the configuration options present or absent in its variants. Selecting and deselecting those configuration options directly influences the functional properties of the system. Apart from functional properties, there are system characteristics that influence the performance (e.g., power demand), safety (e.g., fault probabilities), and security (e.g., susceptibility to attacks) of the system, called Non-Functional Properties (NFPs). Knowledge of NFPs is crucial for evaluating a system’s feasibility, usability, and resource demands. Although variability influences these characteristics, NFPs do not compose linearly for every selected feature. Feature interactions can increase the overall NFP values through (potentially exponential) amplification or decrease them through mitigation effects. In this paper, we propose an automated machine learning (AutoML) approach to predict NFP values for new configurations based on previously measured configuration values. Using AutoML, we leverage the advantages of machine learning for predicting NFPs without having to parameterize and fine-tune machine learning models. This approach and the resulting pipeline aim to reduce the complexity of performance prediction for configurable systems. We test the feasibility of our pipeline in a first evaluation on 4 real-world subject systems and discuss cases where AutoML may improve the prediction of NFPs.