In-Time Explainability in Multi-Agent Systems: Challenges, Opportunities, and Roadmap

Abstract

In the race for automation, distributed systems are required to perform increasingly complex reasoning to deal with dynamic tasks, often not controlled by humans. On the one hand, systems dealing with strict-timing constraints in safety-critical applications mainly focused on predictability, leaving little room for complex planning and decision-making processes. Indeed, real-time techniques are very efficient in predetermined, constrained, and controlled scenarios. Nevertheless, they lack the necessary flexibility to operate in evolving settings, where the software needs to adapt to the changes of the environment. On the other hand, Intelligent Systems (IS) increasingly adopted Machine Learning (ML) techniques (e.g., subsymbolic predictors such as Neural Networks). The seminal application of those IS started in zero-risk domains producing revolutionary results. However, the ever-increasing exploitation of ML-based approaches generated opaque systems, which are nowadays no longer socially acceptable—calling for eXplainable AI (XAI). Such a problem is exacerbated when IS tend to approach safety-critical scenarios. This paper highlights the need for on-time explainability. In particular, it proposes to embrace the Real-Time Beliefs Desires Intentions (RT-BDI) framework as an enabler of eXplainable Multi-Agent Systems (XMAS) in time-critical XAI.