Conditional Abstraction Trees for Sample-Efficient Reinforcement Learning - Oral Session 4

Explore a groundbreaking approach to reinforcement learning in this 25-minute oral presentation from the Uncertainty in Artificial Intelligence conference. Delve into the novel top-down method for constructing state abstractions while simultaneously conducting reinforcement learning. Discover how this domain-independent technique dynamically computes abstractions based on the dispersion of temporal difference errors in abstract states as the agent learns and acts. Examine the extensive empirical evaluation across multiple domains and problems, showcasing how this approach automatically learns semantically rich abstractions finely-tuned to specific problems. Gain insights into the significant improvements in sample efficiency and overall performance compared to existing methods. Learn how this innovative technique addresses the challenge of learning problem abstractions and solutions concurrently in real-world scenarios, potentially revolutionizing the field of reinforcement learning.

UAI 2023 Oral Session 4: Conditional Abstraction Trees for Sample Efficient Reinforcement Learning