Pseudo neural networks via analytic programming with direct coding of constant estimation

Abstract

This research deals with a novel approach to classification - pseudo neural networks (PNN). This technique was inspired in classical artificial neural networks (ANN), where a relation between inputs and outputs is based on the mathematical transfer functions and optimised numerical weights. Compared to ANN, the whole structure in PNN, i.e. the relation between inputs and output(s), is fully synthesised by evolutionary symbolic regression tool - analytic programming. Compared to previous synthesised models, the PNN in this paper were synthesised via a new approach to constant estimation inside the analytic programming - direct coding. Iris data was used for the experiments and PNN were used for the synthesis of a complex classifier for more classes. For experimentation, Differential Evolution (de/rand/1/bin) for optimisation in analytic programming (AP) was used. © ECMS Lars Nolle, Alexandra Burger, Christoph Tholen, Jens Werner, Jens Wellhausen