AURA - Abstraction for Understandability of Reasoning in AI

The recent years witnessed the growth of Artificial Intelligence (AI) research with the designed AI agents be- coming more and more skillful. The increase in the use of such agents makes it crucial for humans to have an understanding of their behavior. However, these agents are usually designed with highly complex structures which makes such an understanding difficult. Knowledge Representation and Reasoning (KRR) is a field of AI where the researchers have been investigating over decades flexible and powerful techniques to expressively represent knowl- edge and empower the agents with reasoning capabilities. Such symbolic and rule-based representations are the key to have more transparency in AI. However it is still challenging for humans to get to the core of the behavior with such representations when they become more complex or contain many distracting details. Towards tackling this problem, this project proposes to make use of abstraction, which is a method that is unwit- tingly used by humans for reasoning to simplify the problem at hand to one that is easier to deal with and to understand. This ability of humans makes it possible to distinguish the relevant details and obtain a high-level understanding. With this project, we will explore ways to employ such human-inspired abstractions to obtain the key elements of rule-based programs that reflect relevant details only and allow for program analysis at the abstract level. We will establish a theoretical foundation for determining good abstractions in terms of distinguishing the key elements for reasoning which will be useful for human-understandability. We will engage in the challenge of automatically computing ab- stractions that capture the essence of reasoning to aid not only in explainability of decision-making but also in general understandability of symbolic and rule-based programs. The investigations in this project will contribute to bringing a KRR perspective to explainability of AI systems, by employing more automated and human-inspired concepts into problem solving.