Replacing generality with coverage for improved learning classifier systems

Measures of Performance of Learning Classifier Systems (LCSs), especially when competing systems both have the same (often 100%) accuracy on the test data, traditionally involve the concept of generality. Put simply, one set of rules is usually favoured over another, all else being equal, if its rules are more general. The notion of generality is also used within LCS algorithms during learning, particularly the XCS system, to promote the generation of a highly general ruleset at the end of the run. In this paper we criticise the standard way that generality is measured, finding that it does not necessarily correlate with high coverage of the space of possible inputs. That is, a highly general result may be preferred because high generality should mean that it can respond to a wider range of inputs, however we indicate that this is not necessarily so. We describe a coverage measure which more usefully reflects this responsiveness, indicating how this can be calculated for single classifiers (rules) or sets of classifiers. We discuss and analyse this measure, showing how it can be used within an LCS algorithm and as a measure of performance.