Section: New Results

Dynamic Decision Networks for Self-Adaptive Systems

Participants : Amel Belaggoun, Nelly Bencomo, Valérie Issarny, Peter Sawyer.

Different modeling techniques have been used to model requirements and decision-making of self-adaptive systems [25] . Important successful techniques based on goal models have been prolific in supporting decision-making according to partial and total fulfillment of functional (goals) and non-functional requirements (softgoals). The final decision about what strategy to use is based on a utility function that takes into account the weighted sum of the different effects of the non-functional requirements. Such solutions have been used both at design and run time including our own solutions using runtime goal models. Different modeling techniques have been used to model requirements and decision-making of self-adaptive systems [25] . Important successful techniques based on goal models have been prolific in supporting decision-making according to partial and total fulfillment of functional (goals) and non-functional requirements (softgoals). The final decision about what strategy to use is based on a utility function that takes into account the weighted sum of the different effects of the non-functional requirements. Such solutions have been used both at design- and run-time including our own solutions using runtime goal models.

We have enriched the decision-making supported by goal models with the use of Bayesian Dynamic Decision Networks (DDNs) [12] . Our novel approach supports reasoning about partial satisfaction of soft-goals using probabilities and uses machine learning. When using DDNs, we introduce new ways to tackle uncertainty based on probabilities that can be updated based on runtime evidence. We have reported the results of the application of the approach on two different cases, one of them being the case of dynamic reconfiguration of a remote data mirroring network that must spread data among servers while minimizing costs and loss of data. Our early results suggest the decision-making process of self-adaptive systems can be improved by using DDNs.

This work has been developed under the umbrella of the Marie Curie Project Requirements@run.time (§  8.2.1.4 ). The main results achieved during the year 2013 are:

• A Bayesian-based technique to support the decision making of self-adaptive systems [14] . DDN-based approaches adopt probabilistic methods (i.e., Bayesian methods) and decision theory to assess the consequences of uncertainty. Using the approach, suitable choices to satisfice functional requirements of the system are identified from a range of alternative decisions and their expected utilities. Satisficement of NFRs is modeled using conditional probabilities given the design decisions. Preferences over decisions are modeled using weights associated with pairs of design alternatives and NFRs, and used when computing the expected utilities of the architectural design alternatives. The decision taken by the DDN is that with the highest expected utility. The approach offers the benefits of machine learning.

• A formal Bayesian definition of surprise as the basis for quantitative analysis to measure degrees of uncertainty and deviation of self-adaptive systems from normal behavior [13] . Specifically, a Bayesian surprise quantifies how new evidence affects assumptions of the world (properties in the models). A “surprising" event may provoke a large divergence between the beliefs distributions prior and posterior to that event. As such and depending on how big or small this divergence is, the running system may decide to either: (i) dynamically adapt accordingly, or (ii) temporarily avoid any action of adaptation and flag up the fact that a potential abnormal situation has been found. While doing (ii) we are offering a specific implementation of the RELAX language previously developed by Bencomo and her co-authors.