Learning Adaptation Models Under Non-Determinism

Among the approaches that have been proposed to support dynamic adaptation, one can find two distinct techniques that appear to be antagonistic. On the one hand, different adaptation models have been proposed as a mean to capture, in an intelligible way, the valuable knowledge that experts have about the system behavior and how to manage it. However, expert-defined models are typically incomplete, often inaccurate and hard to keep up-to-date as the system evolves. On the other hand, the use of machine learning (ML) has been proposed to find, in a fully automatic manner, the correct adaptation strategies. However, ML requires a large training set of observations, usually collected from long and comprehensive training phases to provide meaningful results. Furthermore, it is not trivial for ML to capture non-determinism in the learnt model, in particular with scenarios where a given adaptation may have different outcomes due to factors that have not been taken into account in the original model. In this thesis we present an approach that aims at combining the advantages of static models and machine learning tools as complementary techniques to drive the dynamic adaptation of systems. The approach consists in using the expert’s knowledge to bootstrap the adaptation process and use machine learning to revise, refine, and update the adaptation models at run-time. The revision process is built to take non-determinism into account. The approach has been experimentally validated in a system that performs elastic scaling of RUBiS, a prototype of an auction web application. Palavras Chave Keywords