Methodology for the construction of more efficient artificial neural networks by means of studying and selecting the training set

This article centers upon a less debated, but equally important topic involved in the designing of artificial neural networks (ANNs) which are to be applied to complex problems, such as the specification of architecture or the selection of training parameters. That topic is the preparation of a training set and of a test which are appropriate for the nature of the problem being dealt with. In this paper, we have done this based on a problem of medical nature where the categorizing of patterns depends upon doctors. The problem lies in the great complexity of the example file due to redundancies and some errors. As a result, patterns cannot be identified with complete certainty, and even more, there may exist multiple input and output patterns that fail to be identified individually by the doctors responsible, leading to a loss of resolution to the network.