Canonical form functions as a simple means for genetic programming to evolve human-interpretable functions

摘要

In this paper, we investigate the use of canonical form functions to evolve human-interpretable expressions for symbolic regression problems. The approach is simple to apply, being mostly a grammar that fits into any grammar-based Genetic Programming (GP) system. We demonstrate the approach, dubbed CAFFEINE, in producing highly predictive, interpretable expressions for six circuit modeling problems. We investigate variations of CAFFEINE, including Grammatical Evolution vs. Whigham-style, grammar-defined introns, and smooth uniform crossover with smooth point mutation (SUX/SM). The fastest CAFFEINE variant, SUX/SM, is only moderately slower than non-grammatical GP - a reasonable price to pay when the user wants immediately interpretable results.