General randomization advantage beyond the studied setting

Determine whether randomized algorithms are provably more powerful than deterministic algorithms for oracle-based online learning for arbitrary concept classes and in the non-transductive setting.

Background

The paper resolves the deterministic–randomized separation for thresholds on an unknown order in the transductive model with a consistency-type ERM oracle. It explicitly leaves unresolved whether an analogous separation holds for arbitrary concept classes and outside the transductive protocol, which is identified as the general form of the question posed by Attias, Hanneke, and Ramaswami.