From a Series of (Un)fortunate Events to Global Explainability of Runtime Model-Based Self-Adaptive Systems

Self-adaptive systems (SAS) increasingly use AI-based approaches for their flexible decision-making, which often appear to users as "black boxes". These systems can exhibit unexpected and surprising behaviours that may violate imposed constraints. Runtime models (RTMs) have been used for SAS management in order to provide capabilities needed to explain reasons why the system present the current emergent behaviour. Existing work on explanations derived from RTMs have focused on justifying why the system has presented a specific behaviour at a given time. Nevertheless, we argue that a more general scope is required for understanding the entire evolution of the system, rather than understanding the behaviour for a given instance or situation. From the point of view of Explainable AI (XAI), the latter type of explanations are called global explanations, whereas understanding a single decision refers to local explanations. Global explanations tend to promote trust on the system in question, while local explanations tend to promote trust on a specific decision. In this paper, we propose the use of event graph models to construct global explanations from evolving RTMs. Event graphs allow the representation of the system behaviour as a state-time diagram, by indicating the occurrence of events and their relationships. RTMs are incrementally queried to look for situations of interest (i.e. events), using Complex Event Processing (CEP) in order to analyze and correlate real-time events and therefore, derive conclusions. The approach is applied to a AI-enhanced SAS in the domain of mobile communications. The encouraging results show that event graphs allow the system to present a summarised overview of the system's behaviour, promoting understandability and trustworthiness.