The paper will describe an analysis of a set of real industrial data on the relationship between the cost and size of chemical process pressure vessels. The measures of size include: height, diameter, wall thickness, type, orientation and material of construction, together with the number and size of nozzles. The purpose of the investigation is to provide the company with methods of rapidly estimating the approximate final cost of the item when no detailed design work has been attempted. The approaches to be compared include multi linear regression (MLR), neural networks (NN), fuzzy matching, rational functions and other non-linear models.
MLR was used to give a base point from which to judge the other methods. We have shown that neural networks, having a particularly simple structure, can provide good estimates from a minimum amount of information and the technique of fuzzy matching can yield reasonable accuracies. However, because we had a limited amount of data we were keen to reduce the number of fitted parameters to a minimum. In this regard, the rational and new non-linear forms had much to recommend them. With a small original data set, the non-linear models seemed to be the most attractive method, but if a larger resource was available for analysis then the NN and fuzzy matching ideas had much merit.