Online Learning and Bandits


Follow-the-Leader

WIP (basic baseline) (oscillates in adversarial or worse-case environments)

Follow-the-Regularized-Leader:

WIP (stabilizes previous) (assumes full information)

Exponential-weight Algorithm for Exploration and Exploitation:

WIP (introduce the Bandit setting and the Exploration vs. Exploitation tradeoff) (great for adversarial, explores too aggressively in stochastic settings)

Upper Confidence Bound 1:

WIP (introduces optimism in the face of uncertainty) (context-free, assume static set of arms)

Linear Upper Confidence Bound:

WIP (introduces Contextual Bandit Setting)