Abstract
Machine learning (ML) has shown its potential in extending the ability of self-adaptive systems to deal with unknowns. To date, there have been several approaches to applying ML in different stages of the adaptation loop.
However, the systematic inclusion of ML in the architecture of self-adaptive applications is still an objective that has not been very elaborated yet. In this paper, we show one approach to address this by introducing the concept of estimators in an architecture of a self-adaptive system.
The estimator serves to provide predictions on future and currently unobservable values via ML. As a proof of concept, we show how estimators are employed in ML-DEECo-a dedicated ML-enabled component model for adaptive component architectures.
It is based on our DEECo component model, which features autonomic components and dynamic component coalitions (ensembles). It makes it possible to specify ML-based adaptation already at the level of the component-based application architecture (i.e., at the model level) without having to explicitly deal with the intricacies of the adaptation loop.
As part of the evaluation, we provide an open-source implementation of ML-DEECo run-time framework in Python.