Integrated Planning and Reinforcement Learning for Compositional Domains
Harsha Kokel, Research Scientist
IBM Research

Many real-world domains involving sequential decision-making exhibit a compositional nature. Such domains possess an inherent hierarchy that allows decision-makers to abstract away certain details and focus on making high-level decisions using abstractions. In this talk, I will present an overview of some of our recent efforts in leveraging a combination of planning and reinforcement learning for such domains. First, I will present our approach to constructing task-specific abstractions from the influence information in the domain. Second, I will talk about using that in a framework for learning efficient and generalizable agents. Finally, I will discuss the difference between the action spaces of two sequential-decision-making formulations (MDP & PDDL task) and propose an approach to bridge that gap.

Harsha Kokel is a Research Scientist at IBM Research. Her research focuses on efficient knowledge-guided learning in structured, relational domains. She is interested in sequential decision-making problems and exploring the combination of planning and reinforcement learning.  She earned her Ph.D. at the University of Texas at Dallas. Her research has been published in the top AI/ML conferences including AAAI, IJCAI, and NeurIPS. She is currently co-organizing workshops on bridging the gap between planning and RL; at ICAPS & IJCAI 2023. She also serves as an assistant electronic publishing editor for JAIR.

----
AI Seminar Important Reminders:
-> For graduate students in the AI program, attendance is strongly encouraged