Overview
Explore a comprehensive explanation of the paper "Dynamics-Aware Unsupervised Discovery of Skills" in this 50-minute video. Dive into the world of reinforcement learning as the presenter breaks down a novel framework capable of discovering low-level skills without rewards and composing them to reach specified goals without additional learning. Begin with the motivation behind this approach, then progress through a high-level overview and comparison of model-based and model-free reinforcement learning. Examine the concept of skills, mutual information objectives, and their decomposition. Delve into the unsupervised skill discovery algorithm and learn about planning in skill space. Be prepared for a math-heavy discussion that covers complex topics in artificial intelligence and machine learning. Gain insights into this innovative approach that combines model-based and model-free learning to create more versatile and adaptable AI systems.
Syllabus
- Motivation
- High-Level Overview
- Model-Based vs Model-Free Reinforcement Learning
- Skills
- Mutual Information Objective
- Decomposition of the Objective
- Unsupervised Skill Discovery Algorithm
- Planning in Skill Space
- Conclusion
Taught by
Yannic Kilcher