Explore optimal anytime regret with two experts in this 21-minute IEEE conference talk. Delve into prediction with expert advice, regret bounds for expert learning, and key techniques. Examine the presenters' results and follow a comprehensive proof overview. Understand simplifications for two experts and how Algorithm A depends on time and gap. Investigate discrete and continuous regret tasks, focusing on designing pt and p(t,g). Learn about the Backward Heat Equation and its relation to regret tasks. Discover the initialization process of Algorithm A and draw insightful conclusions from this in-depth exploration of expert learning and regret optimization.
Overview
Syllabus
Optimal anytime regret with two experts
Prediction with expert advice
Regret bounds for expert learning
Techniques
Our result
Proof Overview
Simplifications for two experts Alg. A depends only on time t and "gap" g
Discrete Regret Summary Task: Design pt.
Continuous Regret Task: Design p(t.g)
Backward Heat Equation and Regret Task: Design pt. g
The Backward Heat Equation
Back to discrete From Ito's Formula...
The algorithm Algorithm A initializes go = 0
Conclusion
Taught by
IEEE FOCS: Foundations of Computer Science
