Symbolic Regression for Knowledge Discovery

This work describes an approach for data analysis based on symbolic regression and genetic programming, that produces an overall view of the dependencies of all variables of a system. The identified dependencies are represented in form of a variable interaction network.
In the first part of this work, this approach is described in detail. Important issues are the prevention of bloat and overfitting, the simplification of models, and the identification of relevant input variables. In this context, different methods for bloat control are presented and compared. In addition, a novel way to detect and reduce overfitting is presented and analyzed.
The second part of this work demonstrates how comprehensive symbolic regression can be applied for analysis of real-world systems. Variable interaction networks for a blast furnace process and an industrial chemical process are presented and discussed. Additionally, the same approach is also applied on an economic data set to identify macro-economic dependencies.