It is necessary to determine the size of a hierarchical neural network before we make the neural network learn some data. Moreover, the structure of a network has a large influence on the basic abilities to realize mapping between input and output and produce the generalization power. Generally, a network structure is determined by trial and error and on the basis of the experience and knowledge of the designers.
In this paper, we propose a Structural Learning Algorithm for a network which begins and carries out learning from a bigger network, which reduces redundant links according to fuzzy reasoning, and removes units based on their contribution. Stress is placed on the following points: It should not be necessary to determine the detailed structure of the network in advance, the proposed algorithm can reduce redundant links and units in the learning process, it enables us to build as compact a structure as possible, and also enables mapping between input and output. Furthermore, the proposed method can reduce the learning time which has been considered an unavoidable cost in the construction of a neural network, and sufficiently decrease its average of squared errors.