0-1 Quadratic Knapsack Problems: An Exact Approach Based on a $t$-Linearization

• Published in: SIAM journal on optimization Vol. 22; no. 4; pp. 1449 - 1468
• Main Authors: Rodrigues, C. D., Quadri, D., Michelon, P., Gueye, S.
• Format: Journal Article
• Language: English
• Published: Philadelphia Society for Industrial and Applied Mathematics 01.01.2012
• Subjects: Algorithms; Computer Science; Knapsack problem; Variables; 0-1 quadratic programming; branch-and-bound AMS subject classifications 90C20; 93B18; 90C57; linearization
• ISSN: 1052-6234; 1095-7189
• DOI: 10.1137/110820762

This paper presents an exact solution method based on a new linearization scheme for the 0-1 quadratic knapsack problem, which consists of maximizing a quadratic pseudo-Boolean function with nonnegative coefficients subject to a linear capacity constraint. Contrasting with traditional linearization schemes, our approach adds only one extra variable. The suggested linearization framework provides a tight upper bound, which is used in a branch-and-bound scheme. This upper bound is numerically compared with that of [A. Billionnet, A. Faye, and E. Soutif, European J. Oper. Res., 112 (1999), pp. 664--672], and our branch-and-bound scheme with the exact algorithm of [W. D. Pisinger, A. B. Rasmussen, and R. Sandvik, INFORMS J. Comput., 19 (2007), pp. 280--290]. The experiments show that our upper bound is quite competitive (less than $1\%$ from the optimum). In addition, the proposed branch-and-bound clearly outperforms the algorithm developed by Pisinger et al. for low density instances ($25\%$) for all instances up to $400$ variables. [PUBLICATION ABSTRACT]