Year: 2017
Type: dissertation
Abstract: The strictly convex integer quadratically constrained problem (IQCP) is an important class of optimization problem that arises in numerous applications both in theoretical science and industry. Due to its NP-hard nature, common exact methods are typically tree search algorithms with various techniques for pruning sub-trees while ensuring that at least one optimal solution is retained. In this dissertation, we exploit the similarities among three tree search algorithms, discrete ellipsoid-based search (DEBS), mixed integer programming (MIP), and constraint programming (CP) and integrate their techniques to develop efficient hybrid algorithms for solving IQCPs. We empirically demonstrate the superior performance of our hybrid algorithms in comparison to the previous best known exact methods in the literature. Our novel hybrid algorithms achieve the state-of-the-art on a variety of important classes of the IQCP, such as the binary quadratic programming problem, the exact quadratic knapsack problem, the variance minimization problem, and the integer least squares problems, which all have significant influence in the research fields of operations research (OR) and communications. In addition, we propose a novel global constraint in CP that infers domain reductions based on strictly convex quadratic constraints. Our new global constraint, together with the complementary branching heuristics, bring CP within an order of magnitude of the best-known techniques for the IQCPs tested, and achieve the state of the art when compared with general solvers such as CPLEX on some problem domains. Our new global constraint and branching heuristics fit naturally in the ``model and solve" framework of CP and provide much more flexibility than the hand-crafted, specialized algorithms, allowing practitioners to easily adapt our techniques. For the first time our study brings together the research of OR, CP and communications communities on IQCPs. We believe that our success in connecting different research fields can provide new insights into IQCPs and encourage future innovation.
Source: TSpace (University of Toronto)
Author: Wen-Yang Ku
Language: English
Subfield: Numerical Analysis
Field: Mathematics
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W2767377261
