Learning non-deterministic impact models for adaptation

Many adaptive systems react to variations in their environment by changing their configuration. Often, they make the adaptation decisions based on some knowledge about how the reconfiguration actions impact the key performance indicators. However, the outcome of these actions is typically affected by uncertainty. Adaptation actions have non-deterministic impacts, potentially leading to multiple outcomes. When this uncertainty is not captured explicitly in the models that guide adaptation, decisions may turn out ineffective or even harmful to the system. Also critical is the need for these models to be interpretable to the human operators that are accountable for the system. However, accurate impact models for actions that result in non-deterministic outcomes are very difficult to obtain and existing techniques that support the automatic generation of these models, mainly based on machine learning, are limited in the way they learn non-determinism.

In this paper, we propose a method to learn human-readable models that capture non-deterministic impacts explicitly. Additionally, we discuss how to exploit expert's knowledge to bootstrap the adaptation process as well as how to use the learned impacts to revise models defined offline. We motivate our work on the adaptation of applications in the cloud, typically affected by hardware heterogeneity and resource contention. To validate our approach we use a prototype based on the RUBiS auction application.