Tags, Interaction Patterns, and the Evolution of Cooperation

Abstract

There are several key ideas that appear in almost all of John Holland's writings on artificial and natural complex adaptive systems: internal models, default hierarchies, genetic (evolutionary) algorithms, and recombination of building blocks. One other mechanism, which is linked to all of those, is tag-based interaction. Perhaps the first use of tag-based interaction (though it was not so named) can be found in Holland's "broadcast system," [26] a formal specification of an architecture suitable for modeling adaptation of open-ended, parallel processes. Tag-based interaction mechanisms next played a key role in classifier systems [30, 32]. In classifier systems, a tag acts as a kind of "address" of one or more classifier rules (productions), enabling rules to send messages to selected sets of rules, and allowing rules to select which messages they will respond to. Thus, tags provide a way to structure computations, making it possible to prove that classifier systems are computationally complete [18], to various neural network architectures [8, 55] and even to abstract models of immune systems [17]. Tags also are used to form coupled chains of classifiers, to construct subroutinelike structures, and to allow Holland's Bucket Brigade algorithm to efficiently allocate credit to "stage setting" rules [9, 30, 50]. Holland has also described how tagged classifiers might be used to form default hierarchies and other more complex internal models [28, 30, 33, 46]. More generally, Holland has emphasized the key role that tag-based interaction mechanisms have in almost all complex adaptive systems (CAS), i.e., systems composed of limited capability agents who interact to generate systemlevel behavior [31]. In the context of CAS, tags are arbitrary properties or traits of agents which are visible to other agents, and which agents can detect and use to condition reactions to other tag-carrying agents. Tags can be agent features, such as surface markings, or they can be agent behaviors, from behavioral routines in animals to more complex behaviors of humans, e.g., wearing particular clothes, carrying flags, or following religious customs [3, 31, 53]. Since agents can have different tags, and since arbitrary tags can come to be associated with particular types of agents (with their own interaction and behavioral patterns), tags can take on "meanings" by virtue of the types of agents who display each particular tag, i.e., as a result of the other behavioral traits those agents tend to have.