Symbolic Regression Is Not Enough: It Takes a Village to Raise a Model

Abstract

From a real-world perspective, good enough has beenachieved in the core representations and evolutionarystrategies of genetic programming assumingstate-of-the-art algorithms and implementations arebeing used. What is needed for industrial symbolicregression are tools to (a) explore and refine thedata, (b) explore the developed model space and extractinsight and guidance from the available sample of theinfinite possibilities of model forms and (c) identifyappropriate models for deployment as predictors,emulators, etc. This chapter focuses on the approachesused in DataModeler to address the modelling lifecycle. A special focus in this chapter is theidentification of driving variables and meta variables.Exploiting the diversity of search paths followedduring independent evolutions and, then, looking at thedistributions of variables and metavariable usage alsoprovides an opportunity to gather key insights. Thegoal in this framework, however, is not to replace themodeller but, rather, to augment the inclusion ofcontext and collection of insight by removingmechanistic requirements and facilitating the abilityto think. We believe that the net result is higherquality and more robust models.