Toward curious learning classifier systems

This paper proposes a novel approach to enhance the rather reactive knowledge generation process in Learning Classifier Systems (LCS) toward a more proactive means. We describe how concepts from the domain of Active Learning can be adapted to XCS's algorithmic structure to introduce 'curiosity'. The overall goal is to allow LCSs to build up new knowledge before it is actually requested during the online learning process. We deem such a methodology meaningful in scenarios where data samples are distributed non-uniformly and partially sparse over the input space. Such data imbalances result in gaps within the knowledge base, i.e. an LCS population. We underpin the general potential of our approaches by presenting preliminary results on a realistic data set from the domain of medical diagnosis as well as on a novel toy problem.