A theory for proactive learning of new knowledge and new abilities in AI systems

Research in AI system autonomy has so far focused on solving the question how to act, given a goal by the human. The emerging field of proactivity investigates what to do and when, considering current and future state development, with no goal given. Proactivity can be defined as self-initiated, anticipatory action. This project aims to go yet another step further. Its purpose is to investigate how AI systems proactively can decide what new knowledge and what new ability to learn and when. The system can do so by reasoning on its own and others’ knowledge, on which knowledge to learn for learning a new ability, and which ability to learn for learning a new knowledge. The significance of this research roots in the growing number of human-AI systems. Proactivity is a human trait and humans expect it from their collaborators. Hence, proactive AI systems can facilitate the collaboration with humans. Next level proactive systems, which this project aims to realize, can therefore make human-AI system collaboration even more efficient. We develop a general formal theory and computational methods enabling reasoning about knowledge, ability and learning. We implement our proactively learning systems in simulation and in physical robot systems, which we evaluate in experiments including humans.