What Is a Learning Classifier System?

Abstract

Abstract.We asked "What is a Learning Classifier System" to some of the best-known researchers in the field. These are their answers. John H. HollandClassifier systems are intended as a framework that uses genetic algorithms to study learning in condition/action, rule-based systems. They simplify the "broadcast language" introduced in [26] by (i) eliminating the ability of rules to generate other rules, and (ii) by simplifying the specification of conditions and actions. They were introduced in [27] and were later revised to the current "standard" form in [28]. [31] gives a comprehensive description of this "standard" form, with examples. There are, however, many significant variants (e.g., Booker [9, this volume] In defining classifier systems I adopted the common view that the state of the environment is conveyed to the system via a set of detectors (e.g., rods and cones in a retina). The outputs of the detectors are treated as standardized packets of information, messages. Messages are used for internal processing as well, and some messages, by directing the system's effectors (e.g., its muscles), determine the system's actions upon its environment. There is a further environmental interaction that is critical to the learning process: the environment must, in certain situations, provide the system with some measure of its performance. Here, as earlier [26], I will use the term payoff as the general term for this measure.

Learning Classifier Systems: A Complete Introduction, Review, and Roadmap

“…Together, [45,165] represent two decadelong consecutive summaries of current systems, unsolved problems and future challenges. For comparative system discussions see [102,192,263,264]. For a detailed summary of LCS community resources as of (2002) see [265].…”

ELF-Miner: using structural knowledge and data mining methods to detect new (Linux) malicious executables

“…Similarly, we have used four well known bio-inspired genetic machine learning algorithms: (1) cAntMiner [20] is ant colony optimization based classifier, (2) XCS is a Michigan style learning classifier system [31], (3) UCS is optimized for supervised learning environments [10], and (4) GAssist ADI is a Pittsburgh style learning classifier system [9]. We skip details of classifiers in this paper for brevity, but an interested reader may consult [15] for their description.…”

On the appropriateness of evolutionary rule learning algorithms for malware detection

“…For a given attribute X and a class attribute Y , the uncertainty is given by their respective entropies H(X) and H(Y ). Then the information gain of X with respect to Y is given by I(Y ; X), where 1 The terms accuracy and classification accuracy are interchangeably used in this paper. …”

“…But evolutionary rule learners have received significantly more attention compared with other LCS. They can be further subdivided into Michigan-style, Pittsburghstyle, anticipatory, and iterative rule learners [1]. Most of the recent work in LCSs is focused on classical Michiganand Pittsburgh-style evolutionary rule learners, which include but are not limited to XCS, UCS, GAssist and GALE.…”