| The combination of applications integrated within a single or multi-computer environment has become a key component in the way many organizations deliver their services and provide support. The increased diversity of applications compounded by rising expectations from users regarding the quality of service performance of these systems means that more and more systems administrators are turning to automated solutions. Policy-based management offers significant benefit to this effect since the use of policies can make it more straightforward to define and modify systems behavior at run-time, through policy manipulation, rather than through re-engineering. Equally important, however, is the need for systems to continuously monitor their own behavior from the use of policies, evaluate that behavior, and adapt, when necessary, to cope with not only the inherent human error but also changes in the configuration of the managed environment.;Keywords: Policy-based Management, autonomic management, reinforcement learning, model adaptation, quality of service management.;This thesis proposes an adaptive policy-driven autonomic management approach to quality of service provisioning, utilizing Reinforcement Learning methodologies to determine how best to use a set of policies to meet performance objectives. We believe that "learning" could offer significant benefits to this effect since it enables systems to learn from past experience, predict future actions, and make appropriate trade-offs when selecting policy actions for resolving quality of service violations and for optimizing resources usage. Contrary to most approaches utilizing some form of learning to guide performance management where changes to the environment (in this case the active set of policies) often mean discarding the old knowledge, this work presents an approach for "re-using" the experience - by transforming a model learned from the use of one set of active policies to a new model when those policies change. Since the strategies for learning and adaptation are dependent only on the structure of the policies, our approach can be utilized in a variety of domains. As such, our work is both unique and essential in developing flexible, adaptive, and portable management solutions. |
