Learning World Graphs to Accelerate Hierarchical Reinforcement Learning

Explore a video lecture on hierarchical reinforcement learning and world graphs. Discover how to divide complex tasks into different levels of coarseness, with top-level agents planning over high-level views and subsequent layers handling more detailed perspectives. Learn about a novel approach that proposes learning important states and their connections as a high-level abstraction. Delve into the two-stage process of jointly training a latent pivotal state model and a curiosity-driven goal-conditioned policy, followed by leveraging the world graph for efficient task solving. Understand how this method can significantly improve performance and efficiency in challenging maze tasks compared to baselines lacking world graph knowledge.

Introduction
Hierarchical reinforcement learning
World Graphs
Goal Condition
Important States
Conclusion