A learning-theory-based training algorithm for variable-structure dynamic neural modeling

Different methods of searching for dynamic neural models with minimum complexity have been proposed. The performance as well as the optimality of such methods highly depend on the way “model complexity” is defined. On the other hand, the learning theory creates a framework to assess the learning properties of models. These properties include the required size of the training samples as well as the statistical confidence over the model. In this paper, we first apply the learning properties of the reciprocal multi-quadratic radial basis function networks to introduce a new measure of complexity, which provides a balance between the training and testing performances of the model. Then, we present a systematic evolutionary programming technique that searches for a neural model of an unknown system with the optimal structure as well as parameters. The performance of the novel evolutionary method is illustrated by a numerical modeling simulation that testifies to the success of the proposed method