Matroid Secretary via Labeling Schemes

TL;DR

This paper introduces a general framework for analyzing greedy algorithms in the Matroid Secretary Problem, establishing new probability-competitiveness bounds for laminar and graphic matroids using a novel language-based approach.

Contribution

The paper presents a new analytical framework based on language construction for greedy algorithms in MSP, improving bounds for laminar and graphic matroids.

Findings

Exact probability-competitiveness of 1 - ln(2) for laminar matroids.

Probability-competitiveness of 0.2693 for simple graphs.

Probability-competitiveness of 0.2504 for general graphs.

Abstract

The Matroid Secretary Problem (MSP) is one of the most prominent settings for online resource allocation and optimal stopping. A decision-maker is presented with a ground set of elements $E$ revealed sequentially and in random order. Upon arrival, an irrevocable decision is made in a take-it-or-leave-it fashion, subject to a feasibility constraint on the set of selected elements captured by a matroid defined over $E$. The decision-maker only has ordinal access to compare the elements, and the goal is to design an algorithm that selects every element of the optimal basis with probability at least $\alpha$ (i.e., $\alpha$-probability-competitive). While the existence of a constant probability-competitive algorithm for MSP remains a major open question, simple greedy policies are at the core of state-of-the-art algorithms for several matroid classes. We introduce a flexible and general algorithmic framework to analyze greedy-like algorithms for MSP based on constructing a language associated with the matroid. Using this language, we establish a lower bound on the probability-competitiveness of the algorithm by studying a corresponding Poisson point process that governs the words' distribution in the language. Using our framework, we break the state-of-the-art guarantee for laminar matroids by settling the probability-competitiveness of the greedy-improving algorithm to be exactly $1-\ln(2) \approx 0.3068$. We also showcase the capabilities of our framework in graphic matroids, to show a probability-competitiveness of $0.2693$ for simple graphs and $0.2504$ for general graphs.