Constant optimization and feature standardization in multiobjective genetic programming

Abstract This paper extends the numerical tuning of tree constants in genetic programming (GP) to multiobjective domain. Using ten real-world benchmark regression datasets and employing Bayesian comparison procedures, we first consider effects feature standardization (without constant tuning) conclude that generally produces lower test errors, but, contrary other recently published work, find much less clear trend for sizes. In addition, – with without observe (1) invariably improves error, (2) usually decreases size. Combined standardization, best error results; sizes, however, are increased. We also examine applying only once at end a conventional GP run which turns out be surprisingly promising. Finally, merits using procedures tune around half evolutionary search alone is superior whereas remaining half, parameter superior. identify number open research questions arise from this work.