Lexicographic parsimony pressure

Abstract

We introduce a technique called lexicographic parsimony pressure, for controlling the significant growth of genetic programming trees during the course of an evolutionary computation run. Lexicographic parsimony pressure modifies selection to prefer smaller trees only when fitnesses are equal (or equal in rank). This technique is simple to implement and is not affected by specific differences in fitness values, but only by their relative ranking. In two experiments we show that lexicographic parsimony pressure reduces tree size while maintaining good fitness values, particularly when coupled with Kozastyle maximum tree depth limits.