A methodology for software cost estimation using machine learning techniques

Abstract

The Department of Defense expends billions of dollars on software development and maintenance annually. Many Department of Defense projects fail to be completed, at large monetary cost to the government, due to the inability of current software cost-estimation techniques to estimate, at an early project stage, the level of effort required for a project to be completed. One reason is that current software cost-estimation models tend to perform poorly when applied outside of narrowly-defined domains. Machine learning offers an alternative approach to the current models. In machine learning, the domain specific data and the computer can be coupled to create an engine for knowledge discovery. Using neural networks, genetic algorithms, and genetic programming along with a published software project data set. several cost estimation models were developed. Testing was conducted using a separate data set. All three techniques showed levels of performance that indicate that each of these techniques can provide software project managers with capabilities that can be used to obtain better software cost estimates