To Bound Search Space and Boost Performance of Learning Classifier Systems: a Rough Set Approach

Learning classifier system is a machine learning technique which combines genetic algorithm with the power of the reinforcement learning paradigm. This rule based system has been inspired by the general principle of Darwinian evolution and cognitive learning. XCS, eXtended Classifier System, is currently considered as state-of-the-art learning classifier systems due to its effectiveness in data analysis and its success in applying to varieties of learning problems. It tries to evolve a rule set as a solution, which can cover the whole problem space effectively. Since XCS does not discriminate between class boundary and non-class boundary regions of input space, learning non-class boundaries with more samples might prevent XCS to learn the class boundary regions precisely. This paper explores XCS from this perspective and provides an elegant approach based on rough set theory to deal with this issue. Here, rough set is a mathematical formalism adapted for XCS, which offers our model, i.e., RSXCS, the ability to analyze the problem space and differentiate certain regions from vague ones. The certain regions are first explored by PSU, as an important component of RSXCS, and the rests are directed to other learning components for further rule discovery. To investigate the advantages of RSXCS, quite a lot of experiments across different UCI data sets are conducted. The results show that our approach presents statistically significant improvement in term of classification accuracy compared with its state-of-the-art rival algorithm, i.e., XCS.