Home > Research > Publications & Outputs > Generalized restless bandits and the knapsack p...

Electronic data

Links

Text available via DOI:

View graph of relations

Generalized restless bandits and the knapsack problem for perishable inventories

Research output: Contribution to journalJournal article

Published
Close
<mark>Journal publication date</mark>05/2014
<mark>Journal</mark>Operations Research
Issue number3
Volume62
Number of pages16
Pages (from-to)696-711
Publication statusPublished
Early online date28/04/14
Original languageEnglish

Abstract

In this paper we introduce the Knapsack Problem for Perishable Inventories concerning the optimal dynamic allocation of a collection of products to a limited knapsack. The motivation for designing such a problem comes from retail revenue management, where different products often have an associated lifetime during which they can only be sold, and the managers can regularly select some products to be allocated to a limited promotion space which is expected to attract more customers than the standard shelves. Another motivation comes from scheduling of requests in modern multi-server data centers so that Quality-of-Service requirements given by completion deadlines are satised. Using the Lagrangian approach we derive an optimal index policy for the Whittle relaxation of the problem in which the knapsack capacity is used only on average. Assuming a certain structure of the optimal policy for the single-inventory control, we prove indexability and derive an efficient, linear-time algorithm for computing the index values. To the best of our knowledge, our paper is the first to provide indexability analysis of a restless bandit with bi-dimensional state (lifetime and inventory level). We illustrate that these index values are numerically close to the true index values when such a structure is not present. We test two index-based heuristics for the original, non-relaxed problem: (1) a conventional index rule, which prescribes to order the products according to their current index values and promote as many products as fit in the knapsack, and (2) a recently proposed index-knapsack heuristic, which employs the index values as a proxy for the price of promotion and proposes to solve a deterministic knapsack problem to select the products. By a systematic computational study we show that the performance of both heuristics is nearly-optimal, and that the index-knapsack heuristic outperforms the conventional index rule.