Solving Single and Multiple Plant Sourcing Problems with a Multidimensional Knapsack Model

This research addresses sourcing decisions and how those decisions can affect the management of a company's assets. The study begins with a single-plant problem, in which one facility chooses, from a list of parts, which parts to bring in-house. The selection is based on maximizing the value of the selected parts, while remaining within the plant's capacity. This problem is defined as the insourcing problem and modeled as a multidimensional knapsack problem (MKP). The insourcing model is extended to address outsourcing and multiple plants. This multi-plant model, also modeled as an MKP, enables the movement of parts from one plant to another and consideration of a company-wide objective function (as opposed to a single-plant objective function as in the insourcing model). The sourcing problem possesses characteristics that distinguish it from the standard MKP. One such characteristic is what we define as multiple attributes. To understand the multiple attribute characteristic, we compare the various dimensions in the multidimensional knapsack problem. A classification is given for an MKP as either having a single attribute (SA) or multiple attributes (MA). Mathematically, the problems of each attribute classification can be modeled in the same way with simply a different interpretation of the knapsack constraints. However, experimentation indicates that the MA-MKP is more difficult to solve than the SA-MKP. For small problems, with 100 variables and 5 constraints, the CPU time required to find the optimal solution for MA-MKP to SA-MKP problems has a ratio of 32:1. To determine effective methods for addressing the MA-MKP, standard mixed integer programming techniques are tested. The results of this testing are that the exact approaches are not successful in dramatically reducing the solution time to the level of the SA problems. However, a simple heuristic that performs very well on the MA-MKP is presented. The heuristic utilizes variations on the benefit-to-cost ratio and strongest surrogate constraints. The results from experimentation for MA-MKP problem sets, generated using the methods for standard MKP test data sets in the literature, are presented and indicate that the heuris-tic performs well and improves with larger problems. The average gap between the heuristic solution and the optimal solution is 1.39% for 200-part problems and is reduced to 0.69% when the size of the problem is increased to 298 parts. Although the MA characteristic reflects the sourcing problem, the actual data used in the experimentation is generated with techniques presented in the literature for standard …