Comparison of Semantic-aware Selection Methods in Genetic Programming

Abstract

This study investigates the performance of several semanticaware selection methods for genetic programming (GP). In particular, we consider methods that do not rely on complete GP semantics (i.e., a tuple of outputs produced by a program for fitness cases (tests)), but on binary outcome vectors that only state whether a given test has been passed by a program or not. This allows us to relate to test-based problems commonly considered in the domain of coevolutionary algorithms and, in prospect, to address a wider range of practical problems, in particular the problems where desired program output is unknown (e.g., evolving GP controllers). The selection methods considered in the paper include implicit fitness sharing (ifs), discovery of derived objectives (doc), lexicase selection (lex), as well as a hybrid of the latter two. These techniques, together with a few variants, are experimentally compared to each other and to conventional GP on a battery of discrete benchmark problems. The outcomes indicate superior performance of lex and ifs, with some variants of doc showing certain potential.

Adaptive Test Selection for Factorization-based Surrogate Fitness in Genetic Programming

“…The SFIMX extensions proposed in this paper corroborate our earlier claims [24,18,26,23,17] that tests not only vary in difficulty, but also that this variability can be exploited to make search more effective. We used this property of test-based problems to shape the probability with which the tests are being drawn for interactions.…”

Lexicase selection of specialists

“…Among the most significant of these variations is epsilon lexicase selection, in which "exactly the lowest error" in the description of the algorithm is replaced with "within epsilon of the lowest error" for a suitably defined epsilon; this has proven to be particularly effective on problems with floating-point errors [16,17]. Additionally, lexicase selection has been effectively used to solve problems in areas such as boolean logic and finite algebras [11,13,18], evolutionary robotics [22], and boolean constraint satisfaction using genetic algorithms [21].…”

Epsilon-Lexicase Selection for Regression

“…Discovery of objectives by clustering (DOC) [12] clusters test cases by population performance, and thereby reduces test cases into a set of objectives for search. Both IFS and DOC were outperformed by lexicase selection on program synthesis and boolean problems in previous studies [6,17]. Other methods attempt to sample a subset of T to reduce computation time or improve performance, such as dynamic subset selection [3], interleaved sampling [4], and co-evolved fitness predictors [22].…”