A Theory of Proactivity: a Step Toward Human-Centered AI

Most of today's AI systems and robots are only capable of reactive behavior: responding to explicit user requests or external events. There is now a growing interest in making those systems proactive, that is, able to act in an anticipatory way and on their own initiative. A few attempts exist in this direction, but we do not have a general theory of what it means for an AI system to be proactive, nor computational tools to realize it. In this project we provide such a theory. We model human-centered proactivity where proactive decisions take into account human preferences and human autonomy. Our theory reasons about the world, predicts its possible evolutions, and enables to proactively decide which actions to execute and when. Our theoretical study is complemented by the development of computational methods to be used in AI systems. We investigate the formal properties of our theory and methods, and empirically validate them with physical robot systems.