Equation Discovery in Databases from Engineering

Abstract

As the quantity of electronically generated engineering data grows rapidly, building computer systems to analyze data automatically and intelligently becomes increasingly important to engineers. The overall process of extracting useable knowledge from electronically stored data is called knowledge discovery in databases. The part of the process where patterns are extracted or models are built is referred to as data mining. This dissertation proposes a data mining method that combines machine learning and regression to help engineers in acquiring knowledge which is preferably expressed as equations. A learning algorithm based on the method has been implemented in the computer system EDDE (Equation Discovery in Databases from Engineering). In addition, to obtain useful models that are understandable to engineers, knowledge specific to the particular problem area is incorporated into EDDE to guide the discovery process. The role of this domain knowledge is investigated. The system EDDE is extensively tested on both synthetic data sets and actual engineering data sets. The tests on synthetic data show that EDDE has some important features, such as not being sensitive to the nwnber of variables in data sets. When compared to other methods (regression tree CART, instances based IBL, multivariate linear regression, model tree MS, neural nets, and combinations of these methods), EDDE generates a smaller size model with lower prediction error. EDDE thus swnmarizes the data more concisely and describes the data better.