How XCS deals with rarities in domains with continuous attributes

Michigan-style learning classifier systems solve problems by evolving distributed subsolutions online. Extracting accurate models for subsolutions which are represented by a low number of examples in the training data set has been identified as a key challenge in LCS, and facetwise analysis has been applied to identify the critical elements for success in unbalanced domains. While models for these elements have been developed for XCS with ternary representation, no study has been performed for XCS with interval-based representation, which is most often used in data mining tasks. This paper therefore takes the original design decomposition and adapts it to the interval-based representation. Theory and experimental evidence indicate that XCS with interval-based representation may fail to approximate concepts scarcely represented in the training data set. To overcome this problem, an online covering operator that introduces new specific genetic material in regions where we suspect that there are rarities is designed. The benefits of the online covering operator are empirically analyzed on a collection of artificial and real-world problems.