Automated Deep Learning and Incremental Retraining-Driven MAPE-K Analyzer Architecture for Intelligent Self-Adaptation

Abstract

Monitoring, Analysis, Planning, and Execution with shared Knowledge (in short, the MAPE-K) build a continuous control loop, which is the cornerstone of self-adaptive systems (SASs). Recently, the integration of Deep Learning (DL) in the Analysis phase has significantly enhanced the decision-making capabilities of dynamic and uncertain SASs. However, when faced with unforeseen conditions or new data features not represented during the training, DL models can generate inaccurate predictions. To address this limitation, we suggest a novel MAPE-K Analyzer architecture that leverages Automated Deep Learning (AutoDL), specifically through the AutoGluon framework, to generate robust DL models whenever the performance of the deployed model degrades. This automated model generation is further complemented by an incremental retraining mechanism to ensure continuous recognition of the evolving data. We evaluate the proposed architecture using datasets produced by DeltaIoT, an IoT simulator tailored for SASs research. Our main contributions are (i) a hybrid approach that combines AutoDL-based model generation with incremental retraining to mitigate performance degradation over time, tailored for the MAPE-K Analyzer component, and (ii) a detailed UML-based architecture that supports this approach. Our experimental results demonstrate that the proposed architecture effectively maintains system performance and robustness under data bias and structural evolution conditions.