Why it matters

Online learning shapes sequential decisions. Understanding shapes ML theory.

Advertisement

The architecture

Predict + observe loss + update.

Regret: total loss vs best fixed action.

Sublinear regret = learning.

Online learning loopPredictat round tObserve lossadversarial or stochasticUpdatefor round t+1Follow-the-regularized-leader + mirror descent + Hedge algorithm all fit framework
Online learning.
Advertisement

How it works end to end

Regret analysis (worst case).

FTRL, mirror descent, Hedge.

Bandits: partial feedback.

OCO: online convex optimization.