Abstract
This paper describes a new neural network modeling technique, named the Self-Designing Neural Network, which incorporates an integrated learning algorithm for structure and weight parameters. Neural networks are frequently applied to various problems such as optimization, control, prediction and data mining because of their nonlinear mapping capability. Although neural networks have wide application, in a multi-layer neural network the establishment of structure parameters such as the number of layers, units and connections between units depends on the experience of the engineer and hence the accuracy of mapping is reduced in many cases. The proposed neural network modeling technique, the concept of which is based on ATR’s neurite network, was developed in order to avoid reliance on engineers. In the proposed method, each neuron grows an axon and a dendrite, and each connection between neurons is automatically constructed in the emergent process. The structure and weight parameters of the neural network in each problem are determined by optimizing the growth process using a genetic algorithm. In this study, the optimization is realized by parallel computation. The validity of the proposed method are investigated by applying the method to Exclusive OR.
