Taming the Monster: A Fast and Simple Algorithm for Contextual Bandits

TL;DR

This paper introduces a fast, simple, and practical algorithm for the contextual bandit problem that leverages an oracle for cost-sensitive classification, achieving optimal regret with minimal oracle calls and demonstrating strong empirical performance.

Contribution

The paper presents a new algorithm that simplifies and accelerates contextual bandit learning by reducing oracle calls while maintaining optimal regret guarantees.

Findings

Achieves statistically optimal regret with  ilde{O}(\sqrt{KT/\log N}) calls

Demonstrates excellent computational and prediction performance in experiments

Most practical approach for general policy classes in contextual bandits

Abstract

We present a new algorithm for the contextual bandit learning problem, where the learner repeatedly takes one of $K$ actions in response to the observed context, and observes the reward only for that chosen action. Our method assumes access to an oracle for solving fully supervised cost-sensitive classification problems and achieves the statistically optimal regret guarantee with only $\tilde{O}(\sqrt{KT/\log N})$ oracle calls across all $T$ rounds, where $N$ is the number of policies in the policy class we compete against. By doing so, we obtain the most practical contextual bandit learning algorithm amongst approaches that work for general policy classes. We further conduct a proof-of-concept experiment which demonstrates the excellent computational and prediction performance of (an online variant of) our algorithm relative to several baselines.