Parametric Cost Analysis Using Neural Networks
Regression, in various forms, is the tool of choice for parametric cost analysis (PCA). Neural networks are a largely unused analysis tool within the cost community. However, neural networks are an effective alternative to all forms of regression for PCA. This paper will illustrate the use of neural networks for PCA and will highlight similarities and differences with respect to the use of regression for PCA. For example, regression provides a linear or transformed linear approximation of parameters for a given set of data. Neural networks provide a purely nonlinear approximation of the parameters for the data. When a nonlinear equation is desired with regression, the nonlinear transformation is assumed and applied to the data prior to regression. Neural networks find the best nonlinear fit with no prior transformation and no assumptions.
This paper will provide a history of the use of neural networks for cost analysis. This paper will also provide an insight into how neural networks process data and how that process and the associated results differ from those of regression. A method of visual analysis that permits the user to see what is happening will be provided.
Also, a method will be provided that permits a valid comparison of the goodness of fit to a given set of data for all forms of estimating, including linear regression, nonlinear transformed linear regression, neural networks, and all other forms of estimating parameters for PCA.
Attendees will leave with the information necessary to learn more about neural networks and how to get started using them for PCA.
Edwin B. Dean
Ed Dean spent fourteen years at the Naval Ordnance Laboratory as a physicist, mathematician, operations research analyst, electronic engineer, software engineer, combat systems engineer, and manager. He spent just over a year at the Naval Supply Systems Command Security Group as a computer security specialist.
He joined NASA at the Langley Research Center in 1983. Until 1990, he was in charge of cost estimating at NASA Langley where he managed, performed, or participated in about sixty five system cost estimates including the Space Exploration Initiative Ninety Day Study (back to the Moon and on to Mars). He then joined the Space Exploration Initiative (SEI) Office at Langley where he focused on cost and logistics aspects of returning to the Moon and the colonization of Mars, as well as technologies for designing systems for cost. From 1994 until March 1997 his research focus was design for competitive advantage and the incorporation of value into multidisciplinary product and process development. From 1994 to 1999 he authored the 500+ page NASA Design for Competitive Advantage website.
He retired from NASA on 2 January 1999 as a senior research engineer in the MultiDisciplinary Optimization Branch of the NASA Langley Research Center with research focus on optimization under uncertainty.
Since retiring from NASA he has performed cost estimating support for PRICE Systems and Valador Inc., and has provided cost data analysis using neural networks for Galorath Inc.
He has published over 70 papers/presentations in simulation, nuclear weapons effects, computers, operations research, cost estimating, quality, engineering management, design for cost, and design for competitive advantage. He is a past Director and past Chairman for the International Society of Parametric Analysts. He was the NASA Langley delegate to the Space Systems Cost Analysis Group where he focused on cost risk.