A Study on Generalization Properties of Artificial Neural Network Using Fahlman and Lebiere's Learning Algorithm

Fahlman and Lebiere's (FL) learning algorithm begins with two-layer network and in course of training, can construct various network architecture. In contrast to the fixed network topology as used in Back propagation (BP) learning algorithm, the inherent flexibility of FL algorithm makes it possible to construct any network topology and this makes FL algorithm more suitable and efficient to build an optimum network for a given task. From this point of view, in this paper, we applied FL algorithm to the same network architecture as BP and compared their generalization properties. The results show that FL algorithm yields excellent saturation of hidden units and consequently has better generalization ability than BP algorithm.