Neural networks represent a methodology widely adopted in different scientific, economic and industrial contexts, thanks to their capability of catching essential system features through direct or indirect observation of system performances during the training phase. In this phase the network is able to build an internal representation of the input/output mapping of the problem under investigation. However this representation is cryptically distributed among the network weights and, in addition, it strongly depends upon the network topology, usually established through a trial and error procedure largely guided by the designer's intuition. For these reasons the network is often seen as a successfully functioning black box or as an expert who denies to communicate the reasons of his/her (correct) judgements and consequent actions. The success of the neural methodology in quite different areas such as pattern recognition, dynamic control etc., is nowadays attracting the scientific community to investigate the problem of the optimal network functioning. In the present paper we consider a feedforward multilayered neural network designed to a reactivity meter based on a simple nuclear reactor model indicate that the Ishikawa structural learning can actually help in the simultaneous achievement of both targets of obtaining an insight in the network design and moreover in the understanding of the internal knowlegde gained by the network during training.
展开▼