Stochastic Knapsack: Semi-Adaptivity Gaps and Improved Approximation

TL;DR

This paper investigates the adaptivity gap in stochastic knapsack problems, providing tighter bounds and a new framework for understanding how limited adaptivity compares to full adaptivity in decision-making.

Contribution

It improves bounds on the adaptivity gap and introduces a semi-adaptivity framework to analyze the effectiveness of limited adaptive queries.

Findings

Tighter bounds on the adaptivity gap in stochastic knapsack.

Quantification of semi-adaptivity gaps for limited queries.

Development of a new analytical approach for adaptive decision trees.

Abstract

In stochastic combinatorial optimization, algorithms differ in their adaptivity: whether or not they query realized randomness and adapt to it. Dean et al. (FOCS '04) formalize the adaptivity gap, which compares the performance of fully adaptive policies to that of non-adaptive ones. We revisit the fundamental Stochastic Knapsack problem of Dean et al., where items have deterministic values and independent stochastic sizes. A policy packs items sequentially, stopping at the first knapsack overflow or before. We focus on the challenging risky variant, in which an overflow forfeits all accumulated value, and study the problem through the lens of semi-adaptivity: We measure the power of $k$ adaptive queries for constant $k$ through the notions of $0$-$k$ semi-adaptivity gap (the gap between $k$-semi-adaptive and non-adaptive policies), and $k$-$n$ semi-adaptivity gap (between fully adaptive and $k$-semi-adaptive policies). Our first contribution is to improve the classic results of Dean et al. by giving tighter upper and lower bounds on the adaptivity gap. Our second contribution is a smoother interpolation between non-adaptive and fully-adaptive policies, with the rationale that when full adaptivity is unrealistic (due to its complexity or query cost), limited adaptivity may be a desirable middle ground. We quantify the $1$-$n$ and $k$-$n$ semi-adaptivity gaps, showing how well $k$ queries approximate the fully-adaptive policy. We complement these bounds by quantifying the $0$-$1$ semi-adaptivity gap, i.e., the improvement from investing in a single query over no adaptivity. As part of our analysis, we develop a 3-step "Simplify-Equalize-Optimize" approach to analyzing adaptive decision trees, with possible applications to the study of semi-adaptivity in additional stochastic combinatorial optimization problems.