A framework for the analysis of manufacturing systems controlled by a production authorization card scheme

The selection of a strategy to control the flow of information and materials in a manufacturing system is an important problem. Buzacott and Shanthikumar (1992) presented a modeling framework, the Production Authorization Card Scheme (PAC), to encompass several different traditional control strategies, such as Kanban, CONWIP, and Base Stock Systems. The challenge they presented was to develop an analysis framework to enable the determination of the best control strategy for a manufacturing system, in terms of the performance measures of such a system, such as average inventory and customer service levels.;While simple systems operating under the PAC scheme may be analyzed analytically under very simplifying assumptions, complex systems do not permit this analysis. Thus, simulation is necessary to study the effects of PAC parameter settings on system performance. In this thesis, we discuss the PACSIM simulation model, which we designed to estimate several performance measures for systems with complexities such as multi-product systems, setup times, cells with multiple servers, and assembly cells.;With the ability to estimate this performance, the problem now becomes the ability to determine the PAC system parameters which provide the best operating strategy in terms of the system performance measures. This would suggest the use of simulation optimization techniques. However, we believe that simulation optimization is inappropriate for this framework. We argue that the correct approach is simulation metamodelling, which is an attempt to approximate the expected value functions of the system performance, with respect to the PAC scheme parameters. Such metamodels provide the ability to-apply deterministic optimization techniques, as well as the ability to explore trade-offs amongst the various performance measures, so that the best control strategy may be selected based on policies and outside decision factors not easily integrated into any optimization approach.;Of the metamodel approaches available, neural networks present many attractive advantages. In this thesis, we demonstrate that feedforward neural networks are a highly practical and feasible approach, even for large numbers of input parameters. One key to training reasonably accurate neural networks is efficient simulation experimental design. We present a sampling strategy to with good space-filling properties, and a set of rules designed to ensure that the PAC parameters included in the input space are feasible and reasonable for future analysis purposes.;We demonstrate the potential of this analysis framework on several manufacturing examples. The systems we selected have from four to ten input parameters. We illustrate with these systems some of the types of analysis permitted by this framework.*.;*This dissertation is a compound document (contains both a paper copy and a CD as part of the dissertation). The CD requires the following system requirements: Windows 95 or higher.