A Learning Classifier System Approach to the Identification of Cellular Automata
Larry Bull and Andrew Adamatzky
This paper presents a novel approach to the programming of automata-based simulation and computation using a machine learning technique. The programming of lattice-based automata for real-world applications is cast as a data mining problem. Our approach to achieving this is to use evolutionary computing and reinforcement learning with performance fed back indicating the predictive accuracy of future behaviour of the given system. The purpose of this work is to develop an approach to learning automata rules that can achieve good performance using data from a variety of kinds of complex systems. Results indicate the approach is effective for both nearest neighbour and non-proximal cellular automata.