Online Sampling and Decision Making with Low Entropy

TL;DR

This paper introduces a near-optimal algorithm for online selection of the top k numbers from n adversarially chosen distributions, achieving a balance between competitive ratio and minimal randomness.

Contribution

It presents a distribution with low entropy that enables a deterministic algorithm to nearly optimally select the largest numbers in an online setting.

Findings

Achieves a competitive ratio of 1 - O(√(log k / k))

Uses entropy Θ(log log n) for the order distribution

Improves previous algorithms with better ratio and lower randomness

Abstract

Consider the problem: we are given $n$ boxes, labeled $\{1,2,\ldots, n\}$ by an adversary, each containing a single number chosen from an unknown distribution; these $n$ distributions are not necessarily identical. We are also given an integer $k \leq n$. We have to choose an order in which we will sequentially open these boxes, and each time we open the next box in this order, we learn the number in the box. Once we reject a number in a box, the box cannot be recalled. Our goal is to accept $k$ of these numbers, without necessarily opening all boxes, such that the accepted numbers are the $k$ largest numbers in the boxes (thus their sum is maximized). A natural approach to solve such problems is to use randomness to sample randomly ordered elements, however, as indicated in several sources, e.g., Turan et al. NIST'15, Bierhorst et al. Nature'18, pure randomness is hard to get in reality. We present an algorithm for this problem, which is provably and simultaneously near-optimal with respect to the achieved competitive ratio and the used amount of randomness. In particular, we construct a distribution on the orders with entropy $\Theta(\log\log n)$ such that a deterministic multiple-threshold algorithm gives a competitive ratio $1-O(\sqrt{\log k/k})$, for $k < \log n/\log \log n$. Our competitive ratio is simultaneously optimal and uses optimal entropy $\Theta(\log\log n)$, improving in three ways the previous best known algorithm, whose competitive ratio is $1 - O(1/k^{1/3}) - o(1)$.