Use of a connection-selection scheme in neural XCSF

XCSF is a modern form of Learning Classifier System (LCS) that has proven successful in a number of problem domains. In this paper we exploit the modular nature of XCSF to include a number of extensions, namely a neural classifier representation, self-adaptive mutation rates and neural constructivism. It is shown that, via constructivism, appropriate internal rule complexity emerges during learning. It is also shown that self-adaptation allows this rule complexity to emerge at a rate controlled by the learner. We evaluate this system on both discrete and continuous-valued maze environments. The main contribution of this work is the implementation of a feature selection derivative (termed connection selection), which is applied to modify network connectivity patterns. We evaluate the effect of connection selection, in terms of both solution size and system performance, on both discrete and continuous-valued environments. © 2010 Springer-Verlag Berlin Heidelberg