Classification with Pseudo Neural Networks Based on Evolutionary Symbolic Regression

This research deals with a novel approach to classification. Classical artificial neural networks, where a relation between inputs and outputs is based on the mathematical transfer functions and optimized numerical weights, was an inspiration for this work. Artificial neural networks need to optimize weights, but the structure and transfer functions are usually set up before the training. There exist some evolutionary approaches, which help to set up the structure or to optimize weights in different ways than standard artificial neural networks do. The proposed method utilizes the symbolic regression for synthesis of a whole structure, i.e. the relation between inputs and output(s). For experimentation, Differential Evolution (DE) and Self Organizing Migrating Algorithm (SOMA) for the main procedure of analytic programming (AP) and DE as an algorithm for meta-evolution were used.