Tracking Significant Changes in Bandit - IFDS 2022

Explore the concept of tracking significant changes in bandit algorithms through this insightful 49-minute talk by Samory Kpotufe from Columbia University, presented at IFDS 2022. Delve into the long-term motivation behind addressing environmental changes in contextual bandits, a common occurrence in practical applications. Examine key contributions, including the intuition that various changes can be safely ignored and the definition of significant phases. Learn about adaptive procedures and the design of random replays. Investigate changes in context distribution and evaluate the informativeness of previous probability distributions. Gain a comprehensive understanding of strategies for dealing with non-stationary environments in bandit algorithms.

Intro
Long Term Motivation
Environmental changes are frequent in practice ©
(Contextual) Bandits
Outline
Changes in Reward Y, distribution
Key Contributions
Intuition: various changes can safely be ignored
Definition (Significant Phases P.)
Adaptive Procedure
Designing Random Replays
Changes in Context X, distribution
Usual (stationary) Strategies
How informative is previous Py for Qx?
Summary