Abductive Design of BDI Agent-based Digital Twins of Organizations

For a Digital Twin - a precise, virtual representation of a physical counterpart - of a human-like system to be faithful and complete, it must appeal to a notion of anthropomorphism (i.e., attributing human behaviour to non-human entities) to imitate (1) the externally visible behaviour and (2) the internal workings of that system. Although the Belief-Desire-Intention (BDI) paradigm was not developed for this purpose, it has been used successfully in human modeling applications. In this sense, we introduce in this thesis the notion of abductive design of BDI agent-based Digital Twins of organizations, which builds on two powerful reasoning disciplines: reverse engineering (to recreate the visible behaviour of the target system) and goal-driven eXplainable Artificial Intelligence (XAI) (for viewing the behaviour of the target system through the lens of BDI agents). Precisely speaking, the overall problem we are trying to address in this thesis is to “Find a BDI agent program that best explains (in the sense of formal abduction) the behaviour of a target system based on its past experiences". To do so, we propose three goal-driven XAI techniques: (1) abductive design of BDI agents, (2) leveraging imperfect explanations and (3) mining belief-based explanations. The resulting approach suggests that using goal-driven XAI to generate Digital Twins of organizations in the form of BDI agents can be effective, even in a setting with limited information about the target system’s behaviour.