Analysis of online signature based learning classifier systems for noisy environments: A feedback control theoretic approach

Abstract

Post training rule set pruning techniques are amongst one of the approaches to improve model comprehensibility in learning classifier systems which commonly suffer from population bloating in real-valued classification tasks. In an earlier work we introduced the term signatures for accurate and maximally general rules evolved by the learning classifier systems. A framework for online extraction of signatures using a supervised classifier system was presented that allowed identification and retrieval of signatures adaptively as soon as they are discovered. This paper focuses on the analysis of theoretical bounds for learning signatures using existing theory and the performance of the proposed algorithm in noisy environments using benchmark synthetic data sets. The empirical results with the noisy data show that the mechanisms introduced to adapt system parameters enable signature extraction algorithm to cope with significant levels of noise.