Falcon : Towards Adaptable and Sustainable Decision-Making in Self-Adaptive Systems

Introducing machine learning in the decision-making enabled self-adaptive systems to perform intelligent adaptation decisions in dynamic environments, by analyzing data and predicting suitable adaptation strategies. However, ML models may produce erroneous predictions when faced with unforeseen conditions or new features that are not encountered during the training phase. In this paper, we propose the Falcon architecture to enhance the sustainability and reliability of Self-Adaptive Systems. In fact, Falcon improves adaptation decisions in response to unexpected environmental changes by dynamically adjusting the system ML models. Hence, it integrates an additional MAPE-K loop that continuously monitors the efficiency of ML predictions based on confidence and response time. Falcon responds to low confidence by adapting the deployed model using transfer learning, to avoid full retraining which can be time-consuming and costly. Fine-tuning is also adopted to maintain sustained accuracy and ensure reliable decision-making in dynamic environments whenever no suitable model is found.