Zoom: https://oregonstate.zoom.us/j/91611213801?pwd=Wm9JSkN1eW84RUpiS2JEd0E5T…

Can we build autonomous agents that learn generalizable knowledge and reliably solve previously unseen problems? In this talk I will present some of my group's recent research on neuro-symbolic learning for sequential decision-making problems that feature long-horizons, sparse rewards and vast differences between training and testing problems. I will discuss methods for learning and using abstractions to learn world models that can be easily transferred to new problem instances, often overshadowing the complexity of training instances. We will see how these methods can be used to invent symbolic vocabularies and learn logic-based world-models for robot task and motion planning without any human annotation or hand-written logic; how learning simple abstractions during Q-learning can vastly improve the performance of RL agents; and finally, how abstractions can help address the emerging problem of AI assessment.