Tag-based modules in genetic programming

Abstract

In this paper we present a new technique for evolving modular programs with genetic programming. The technique is based on the use of "tags" that evolving programs may use to label and later to refer to code fragments. Tags may refer inexactly, permitting the labeling and use of code fragments to co-evolve in an incremental way. The technique can be implemented as a minor modification to an existing, general purpose genetic programming system, and it does not require pre-specification of the module architecture of evolved programs. We demonstrate that tag-based modules readily evolve and that this allows problem solving effort to scale well with problem size. We also show that the tag-based module technique is effective even in complex, non-uniform problem environments for which previous techniques perform poorly. We demonstrate the technique in the context of the stack-based genetic programming system PushGP, but we also briefly discuss ways in which it may be used with other kinds of genetic programming systems.

What else is in an evolved name? Exploring evolvable specificity with SignalGP

“…In Spector et al's original conception of tag-based referencing, as long as a program had at least one tagged module, all referential tags could successfully reference something (Spector et al, 2011b). The tag-based referencing employed by SignalGP, however, can be configured to only match tags whose similarity exceeds a threshold, allowing programs to ignore events by avoiding the use of similar tags.…”

“…Tag-based referencing has built-in flexibility, not requiring tags to exactly match to successfully reference one another. In Spector et al's initial implementation of tag-based referencing (Spector et al, 2011b), referring tags always matched to the most similar receptor tag. Spector et al speculated that tag-based referencing performed well because of this inexactness: any tag-based reference is able to find a referent as long as one exists.…”

“…A 0% similarity threshold is equivalent to what Spector et al (2011b) used in their original demonstration of tag-based referencing; thus, we expected to it perform well. However, are intermediate thresholds just as effective?…”

“…Inspired by John Holland's use of "tags" (Holland, 1987(Holland, , 1990(Holland, , 1993(Holland, , 2006) as a mechanism for matching, binding, and aggregation, Spector et al (Spector et al, 2011b(Spector et al, ,a, 2012 introduced and demonstrated a tag-based naming scheme for GP where tags are used to name and reference program modules. Tags are evolvable labels that are mutable, and the similarity (or dissimilarity) between any two possible tags is quantifiable.…”

“…In Spector et al's original implementation (Spector et al, 2011b), tags were used as an evolvable mechanism to label and later refer to code fragments. At there core, tags provide general-purpose, evolvable specificity -an evolvable way to specify zero or more tagged entities.…”