Application of Symbolic Regression on Blast Furnace and Temper Mill Datasets

摘要

This work concentrates on three different modifications of a genetic programming system for symbolic regression analysis. The coefficient of correlation R2 is used as fitness function instead of the mean squared error and offspring selection is used to ensure a steady improvement of the achieved solutions. Additionally, as the fitness evaluation consumes most of the execution time, the generated solutions are only evaluated on parts of the training data to speed up the whole algorithm. These three algorithmic adaptations are incorporated in the symbolic regression algorithm and their impact is tested on two real world datasets describing a blast furnace and a temper mill process. The effect on the achieved solution quality as well as on the produced models are compared to results generated by a symbolic regression algorithm without the mentioned modifications and the benefits are highlighted.