Transparent, online image pattern classification using a learning classifier system

Abstract

Image pattern classification in computer vision problems is challenging due to large, sparse input spaces with the added demand for generalisation and accuracy of results. The Evolutionary Computation technique of Learning Classifier Systems (LCS) addresses such problems, but has not been applied previously to this domain. Instead, offline, supervised techniques on fixed data sets have been shown to be highly accurate. This paper shows that LCS enable online, reinforcement learning on datasets that may change over time and produce transparent (human readable) classification rules. Further work is needed in domains applicable to offline, supervised learning to achieve benchmark accuracy, but the promising initial results auger well for domains, such as mobile robotics, where compact, accurate and general rules learnt in a graceful manner are required. © 2011 Springer-Verlag.