Cost functions based on different types of distance measurements for pseudo neural network synthesis

Abstract

This research deals with a novel approach to classification. New classifiers are synthesized as a complex structure via evolutionary symbolic computation techniques. Compared to previous research, this paper synthesizes multi-input-multi-output (MIMO) classifiers with different cost function based on distance measurements. An inspiration for this work came from the field of artificial neural networks (ANN). The proposed technique creates a relation between inputs and outputs as a whole structure together with numerical values which could be observed as weights in ANN. Distances used in cost functions were: Manhattan (absolute distances of output vectors), Euclidean, Chebyshev (maximum distance value), Canberra distance, Bray – Curtis. The Analytic Programming (AP) was utilized as the tool of synthesis by means of the evolutionary symbolic regression. For experimentation, Differential Evolution for the main procedure and also for meta-evolution version of analytic programming was used Iris data (a known benchmark for classifiers) was used for testing of the proposed method. © Springer International Publishing Switzerland 2015.