Partially-Connected Artificial Neural Networks Developed by Grammatical Evolution for Pattern Recognition Problems

Abstract

Evolutionary Artificial Neural Networks (EANNs) are a special case of Artificial Neural Networks (ANNs) for which Evolutionary Algorithms (EAs) are used to modify or create them. EANNs adapt their defining components ad hoc for solving a particular problem with little or no intervention of human expert. Grammatical Evolution (GE) is an EA that has been used to indirectly develop ANNs, among other design problems. This is achieved by means of three elements: a Context-Free Grammar (CFG) which includes the ANNs defining components, a search engine that drives the search process and a mapping process. The last component is a heuristic for transforming each GE’s individual from its genotypic form into its phenotypic form (a functional ANN). Several heuristics have been proposed as mapping processes in the literature; each of them may transform a specific individual’s genotypic form into a very different phenotypic form. In this paper, partially-connected ANNs are automatically developed by means of GE. A CFG is proposed to define the topologies, a Genetic Algorithm (GA) is the search engine and three mapping processes are tested for this task; six well-known pattern recognition benchmarks are used to statistically compare them. The aim of this work for using and comparing different mapping process is to analyze them for setting the basis of a generic framework to automatically create ANNs.