PublicationsPapers

The Littlestone dimension of a hypothesis class captures the optimal mistake bound in online learning when the learner is deterministic.

In this work, we define a related parameter, the randomized Littlestone dimension, which captures the optimal mistake bound when the learner is randomized.

Using the new parameter, we prove nearly optimal bounds on prediction with expert advice when the learner is randomized, complementing past work on the deterministic setting.