Antifragility via Online Learning and Monitoring: An IoT Case Study

The introduction of antifragility as a design paradigm for self-adaptive systems is shifting the attention beyond resiliency, which focuses on the ability of systems to recover from failures, towards the ability of such systems to learn from these experiences. Ideally, the manager of an autonomous system should be able to exploit a learning framework and adapt itself alongside the system it is monitoring to deal with unforeseen circumstances. Machine learning, in general, and reinforcement learning, in particular, offer the tools and framework to learn online from experience by extracting models from the collected data. However, learning tools alone are not sufficient: understanding when learning is required is a complementary problem. In fact, learning continuously can turn out to be resource demanding and may lead to faulty models. In this context, monitoring approaches for online learning, a framework of machine learning designed to deal with evolving environments through statistical analysis by predicting when the monitored system has changed, comes in handy. With this paper we offer our view on the problem of implementing antifragility using online learning and monitoring approaches in conjunction with reinforcement learning. Moreover, to support the proposed methodology, we provide a demonstrative implementation applied to the manager of an IoT network and we compare it with alternative learning approaches to showcase its benefits.