Back to the Future—Revisiting OrdinalGP and Trustable Models After a Decade

Abstract

OrdinalGP (2006) [4] embraced a fail-fast philosophy to efficiently model very large data sets. Recently, we realized that it was also effective against small data sets to reward model generalization. ESSENCE (2009) [6] extended the OrdinalGP concept to handle imbalanced data by using the SMITS algorithm to rank data records according to their information content to avoid locking into the behavior of heavily sampled data regions but had the disadvantage of computationally-intensive data conditioning with a corresponding fixed data ranking. With BalancedGP (2019) we shifted to a stochastic sampling to achieve a similar benefit. Trustable models (2007) [3] exploited the diversity of model forms developed during symbolic regression to define ensembles that feature both accurate prediction as well as detection of extrapolation into new regions of parameter space as well as changes in the underlying system. Although the deployed implementation has been effective, the diversity metric used was data-centric so alternatives have been explored to improve the robustness of ensemble definition. This chapter documents our latest thinking, realizations, and benefits of revisiting OrdinalGP and trustable models.