Mining interesting knowledge from data with the XCS classifier system

We apply a version of XCS which exploits a general purpose representation to the problem of mining knowledge from some well-known classification tasks involving synthetic and real-world data. We show that XCS can extract interesting knowledge from data both (i) in terms of predictive accuracy on unseen cases and (ii) in terms of explicit knowledge on the phenomena described in the data. In particular, in synthetic tasks, XCS's predictive accuracy is at least as good as that of more traditional classification algorithms while it can extract rules which give an explicit insight of the data. In real world tasks, XCS outperforms C4.5 in one important medical datasets involving numerical data while it performs quite the same as C4.5 on another real world dataset involving symbolic data.