Breaking the Stereotypical Dogma of Artificial Neural Networks with Cartesian Genetic Programming

Abstract

This chapter presents the work done in the field of Cartesian Genetic Programming evolved Artificial Neural Networks (CGPANN). Three types of CGPANN are presented, the Feed-forward CGPANN (FFCGPAN), Recurrent CGPANN and the CGPANN that has developmental plasticity, also called Plastic CGPANN or PCGPANN. Each of these networks is explained with the help of diagrams. Performance results obtained for a number of benchmark problems using these networks are illustrated with the help of tables. Artificial Neural Networks (ANNs) suffer from the dilemma of how to select complexity of the network for a specific task, what should be the pattern of inter-connectivity, and in case of feedback, what topology will produce the best possible results. Cartesian Genetic Programming (CGP) offers the ability to select not only the desired network complexity but also the inter-connectivity patterns, topology of feedback systems, and above all, decides which input parameters should be weighted more or less and which one to be neglected. In this chapter we discuss how CGP is used to evolve the architecture of Neural Networks for optimum network and characteristics. Don't you want a system that designs everything for you? That helps you select the optimal network, the inter-connectivity, the topology, the complexity, input parameters selection and input sensitivity? If yes, then CGP evolved Artificial Neural Network (CGPANN) and CGP evolved Recurrent Neural Network (CGPRNN) is the answer to your questions.