GECCO'2022 Symbolic Regression Competition: Post-Analysis of the Operon Framework
Bogdan Burlacu
Abstract
Operon is a C++ framework for symbolic regression with the ability to perform local search by optimizing model coefficients using the Levenberg-Marquardt algorithm. This enhancement has proven to be effective in a variety of regression tasks. Operon took part in the Interpretable Symbolic Regression for Data Science hosted at the 2022 Genetic and Evolutionary Computation Conference, where it ranked overall 4th based on criteria of accuracy, simplicity as well as task-specific goals. Although accurate, the returned models were exceedingly complex and ranked poorly in terms of simplicity. In this paper, we investigate the application of the Minimum Description Length (MDL) principle for selecting models with a better compromise between accuracy and complexity from the final Pareto front returned by the algorithm. A new experiment on the synthetic track of the competition highlights the critical role played by model selection in algorithm performance. The MDL-enhanced approach obtains the best overall score and demonstrates excellent results on all synthetic tracks.