An Adaptive Reinforcement Learning Approach to Policy-Driven Autonomic Management

Policies have been explored as a basis for autonomic management. In many cases, there is a need for policy-driven autonomic systems to have the ability to adapt the use of policies based, for example, on past experience, in order to deal with human error or the unpredictability in workload characteristics. This suggests that learning approaches can offer significant potential benefits in providing autonomic systems with the ability to identify preferred uses of existing policies or learn new policies. In this context, we have explored the use of reinforcement learning in adaptive policy-driven autonomic management. A key question is whether a model "learned'' from the use of one set of policies could be applied to another set of "similar'' policies, or whether a new model must be learned from scratch as a result of changes to an active set of policies. In this paper, we illustrate how a reinforcement learning model might be adapted to accommodate such changes.