Monday Computer Science Seminar (22.1.2024)

Qualitative Reasoning in Artificial Intelligence — bridging the gap between machine learning and human reasoning

A fundamental difference between the way conventional methods of Artificial Intelligence (AI) make decisions and the way humans think and reason is that humans reason qualitatively, while AI typically makes decisions based on numerical computations. The gap between the two worlds — the human and the machine — becomes apparent when it comes to exchanging learned knowledge. Traditional numerical models usually consist of a large number of numerical parameters that convey little or no information to a human on why a particular decision or action was taken. On the other hand, it is very difficult for a human to describe their intuitive knowledge of how a particular mechanism works to an AI algorithm in a way that the algorithm can utilize in planning and decision-making. Qualitative Reasoning (QR) is a branch within AI research that focuses on how AI can reason about processes qualitatively, and present the findings in a form that approximates human intuition. I will present the historical origins of Qualitative Reasoning (QR) in AI, its later developments, and the current state of the art. I will focus on the area of agent learning and planning in continuous domains with numerical sensory and actuation systems. We will explore the full cycle of automated abstraction of qualitative representations from numerical observations, the search for symbolic solutions through qualitative reasoning, and the implementation of the found solutions in the original numerical domain. I will explain the foundations of qualitative physics and qualitative simulation, which is the basis for qualitative planning, and thus for predicting possible future behaviors in a symbolic and explainable way. I will present the results of experiments with different robot problems, such as learning to walk, learning to push objects, and learning to swing up and balance a pole.

Domen Šoberl received his PhD in computer science in 2021 from University of Ljubljana, Faculty of Computer and Information Science. He is currently employed as a teaching assistant at UP FAMNIT. His current research interests lie in various areas of artificial intelligence, including deep learning, generative adversarial networks, reinforcement learning, and qualitative reasoning.