The structure of many real-world optimization problems includes minimization of a nonlinear (or quadratic) functional subject to bound and singly linear constraints (in the form of either equality or bilateral inequality) which are commonly called as continuous knapsack problems. Since there are efficient methods to solve large-scale bound constrained nonlinear programs, it is desirable to adapt these methods to solve knapsack problems, while preserving their efficiency and convergence theories. The goal of this paper is to introduce a general framework to extend a box-constrained optimization solver to solve knapsack problems. This framework includes two main ingredients which are O(n) methods; in terms of the computational cost and required memory; for the projection onto the knapsack constrains and the null-space manipulation of the related linear constraint. The main focus of this work is on the extension of Hager-Zhang active set algorithm (SIAM J. Optim. 2006, pp. 526--557). The main reasons for this choice was its promising efficiency in practice as well as its excellent convergence theories (e.g., superlinear local convergence rate without strict complementarity assumption). Moreover, this method does not use any explicit form of second order information and/or solution of linear systems during iteration which makes it an ideal for large-scale problems. The efficient implementation of the method is discussed in details.
R. Tavakoli, Hager-Zhang Active Set Algorithm for Large-Scale Continuous Knapsack Problems, preprint, 2009.