Training Prediction Models for Rule-Based Self-Adaptive Systems

摘要

Architecture-based self-adaptive systems that are rule-based can be steered by predicting changes of the system utility. However, building predictions for these systems is challenging. One of the reasons is that the lack of detailed information about the system performance model makes it difficult to construct an analytic representation of the system utility. We mitigate this problem with a methodology to learn the changes of the system utility without relying on detailed information of the system. We evaluated our methodology over a real system with a range of utility complexities and different machine learning methods trained with real and publicly available failure traces. Our findings suggest that our methodology is applicable to real system failures on dynamic architectures under different configurations.