Self-Optimizing Neural Networks

Abstract

The paper is concentrated on two essential problems: neural networks topology optimization and weights parameters computation that are often solved separately. This paper describes new solution of solving both selected problems together. According to proposed methodology a special kind of multilayer ontogenic neural networks called Self-Optimizing Neural Networks (SONNs) can simultaneously develop its structure for given training data and compute all weights in the deterministic way based on some statistical computations that are incomparably faster then many other training methods. The described network optimization process (both structural and parametrical) finds out a good compromise between a minimal topology able to correctly classify training data and generalization capability of the neural network. The fully automatic self-adapting mechanism of SONN does not use any a priori configuration parameters and is free from different training problems.