21
15

A Near-Optimal Best-of-Both-Worlds Algorithm for Online Learning with Feedback Graphs

Abstract

We consider online learning with feedback graphs, a sequential decision-making framework where the learner's feedback is determined by a directed graph over the action set. We present a computationally efficient algorithm for learning in this framework that simultaneously achieves near-optimal regret bounds in both stochastic and adversarial environments. The bound against oblivious adversaries is O~(αT)\tilde{O} (\sqrt{\alpha T}), where TT is the time horizon and α\alpha is the independence number of the feedback graph. The bound against stochastic environments is O((lnT)2maxSI(G)iSΔi1)O\big( (\ln T)^2 \max_{S\in \mathcal I(G)} \sum_{i \in S} \Delta_i^{-1}\big) where I(G)\mathcal I(G) is the family of all independent sets in a suitably defined undirected version of the graph and Δi\Delta_i are the suboptimality gaps. The algorithm combines ideas from the EXP3++ algorithm for stochastic and adversarial bandits and the EXP3.G algorithm for feedback graphs with a novel exploration scheme. The scheme, which exploits the structure of the graph to reduce exploration, is key to obtain best-of-both-worlds guarantees with feedback graphs. We also extend our algorithm and results to a setting where the feedback graphs are allowed to change over time.

View on arXiv
Comments on this paper