Temporal Models For History-Aware Explainability In Self-Adaptive Systems

Abstract

The complexity of real-world problems requires modern software systems to be able to autonomously adapt and modify their behaviour at runtime to deal with unforeseen internal and external fluctuations and contexts. Consequently, these self-adaptive systems (SAS) can show unexpected and surprising behaviours which stakeholders may not understand or agree with. This may be exacerbated due to the ubiquity and complexity of Artificial Intelligence (AI) techniques which are often considered “black boxes” and are increasingly used by SAS. This thesis explores how synergies between model-driven engineering and runtime monitoring help to enable explanations based on SAS’ historical behaviour with the objective of promoting transparency and understandability in these types of systems. Specifically, this PhD work has studied how the use of runtime models extended with long-term memory can provide the abstraction, analysis and reasoning capabilities needed to support explanations when using AI-based SAS. For this purpose, this work argues that a system should i) offer access and retrieval of historical data about past behaviour, ii) track over time the reasons for its decision making, and iii) be able to convey this knowledge to different stakeholders as part of explanations for justifying its behaviour. Runtime models stored in Temporal Graph Databases, which result in Temporal Models (TMs), are proposed for tracking the decision-making history of SAS to support explanations. The approach enables explainability for interactive diagnosis (i.e. during execution) and forensic analysis (i.e. after the fact) based on the trajectory of the SAS execution. Furthermore, in cases where the resources are limited (e.g., storage capacity or time to response), the proposed architecture also integrates the runtime monitoring technique, complex event processing (CEP). CEP allows detecting matches to event patterns that need to be stored instead of keeping the entire history. The proposed architecture helps developers in gaining insights into SAS while they work on validating and improving their systems.