Messy Coding in the XCS Classifier System for Sequence Labeling

Abstract

The XCS classifier system for sequence labeling (XCS-SL) is an extension of XCS for sequence labeling, a form of time-series classification where every input has a class label. In XCS-SL a classifier condition consists of some sub-conditions which refer back to previous inputs. Each sub-condition is a memory. A condition has n sub-conditions which represent an interval from the current time t0 to a previous time t_ _n . A problem of this representation (called interval coding) is, even if only one input at t_ _n is needed, the condition must consist of n subconditions to refer to it. We introduce a messy coding based condition where each sub-condition messily refers to a single previous time. Unlike the original coding, the set of sub-conditions does not necessarily represent an interval, so it can represent compact conditions. The original XCS-SL evolutionary mechanism cannot be used with messy coding and our main innovation is a novel evolutionary mechanism. Results on a benchmark show that, compared to the original interval coding, messy coding results in a smaller population size and does not require as high a population size limit. However, messy coding requires more training with a high population size limit. On a real world sequence labeling task messy coding evolved a solution that achieved higher accuracy with a smaller population size than the original interval coding