On the complexity of All $\varepsilon$-Best Arms Identification

TL;DR

This paper establishes fundamental lower bounds on the sample complexity for identifying all -optimal arms in Gaussian bandits, proposes an asymptotically optimal strategy, and demonstrates its effectiveness through numerical simulations.

Contribution

It introduces two new lower bounds on sample complexity, designs an asymptotically optimal Track-and-Stop algorithm, and provides an efficient numerical method for solving related convex programs.

Findings

The asymptotic lower bound is tight and guides optimal sampling.

The proposed algorithm outperforms existing methods in simulations.

Sample complexity scales linearly with the number of arms for moderate risks.

Abstract

We consider the question introduced by \cite{Mason2020} of identifying all the $\varepsilon$-optimal arms in a finite stochastic multi-armed bandit with Gaussian rewards. We give two lower bounds on the sample complexity of any algorithm solving the problem with a confidence at least $1-\delta$. The first, unimprovable in the asymptotic regime, motivates the design of a Track-and-Stop strategy whose average sample complexity is asymptotically optimal when the risk $\delta$ goes to zero. Notably, we provide an efficient numerical method to solve the convex max-min program that appears in the lower bound. Our method is based on a complete characterization of the alternative bandit instances that the optimal sampling strategy needs to rule out, thus making our bound tighter than the one provided by \cite{Mason2020}. The second lower bound deals with the regime of high and moderate values of the risk $\delta$, and characterizes the behavior of any algorithm in the initial phase. It emphasizes the linear dependency of the sample complexity in the number of arms. Finally, we report on numerical simulations demonstrating our algorithm's advantage over state-of-the-art methods, even for moderate risks.