A new approach to encoding actions in classifier systems

Extended Abstract The classifier system framework is a general-purpose approach to learning and representation designed to exhibit non-brittle behavior in complex, continually varying environments. Broadly speaking, classifier systems are expected to avoid brittle behavior because they implement processes that build and refine models of the environment. One of the most important of these processes is categorization. As Holland [5] has pointed out (p. 598) " Categorization is the system's major weapon for combating the environment's perpetual novelty. The system must readily generate categories for input messages, and it must be able to generate categories relevant to its internal processes ". Research in classifier systems has focused almost exclusively on finding generalizations for input messages. However , generalizations of actions will also be required in order to build effective models of the environment. This paper introduces a new encoding for actions in classifier rules that lends itself to representing abstract actions. What categorizations of actions are desirable? Two distinct kinds of generalizations appear to be useful [4]: abstractions that generalize over the possible outcomes of an action, and aggregations that group related actions together. Both kinds of generalizations are important for classifier systems, and some previous research has offered preliminary ideas about how to represent them (e.g., the expectation part of rules and action chunks in ACS [8] and the fuzzy combination of action effects in fuzzy classifier systems [1]). One fundamental issue that has not been addressed, though, is that actions are always represented as primitive symbols having no constituent parts. This is problematic in a classi-fier system that depends on a genetic algorithm to discover generalizations by manipulating building blocks. What is needed is an encoding for actions that uses effector settings as the parts of an action, while at the same time permits useful generalizations of some kind. Mechanisms prevalent in natural systems point the way to a solution. Regarding actions themselves, pairs of interacting effectors can provide a powerful set of building blocks for constructing more sophisticated behaviors [2]. As for encodings, the genotype of a heterozygote can be viewed as a reservoir of variability [3] that in some sense is an abstraction of the specific phenotypic traits