The Next Level of Long-Term Agent Autonomy -- Proactively Acquiring Knowledge and Abilities

For an artificial agent operating long-term under real-world conditions it is not enough to be able to act on orders given by the human. Even being able to act proactively (anticipatory, self-initiated) does not suffice. The reason roots in the unrealistic assumption that the proactive agent from the start and always knows everything it needs to know and has all the abilities it requires. We argue that the agent has to be able to proactively learn new knowledge and abilities according to how the dynamic environment evolves. We identify challenges and directions towards proactive learning. Our focus is on formal methods which lend themselves to doing the necessary reasoning but also give suggestions how these might be integrated with machine learning. The ideas envisioned in this paper can advance (M)AS (Multi-Agent Systems) research and have the potential to enhance collaborations of hybrid human-AI systems.