Optimum Experimental Design for Locally Recurrent Networks

Contents

A fundamental problem underlying the training of neural networks is the selection of proper input data to provide good representation of the modelled system behaviour [77, 7]. This problem comprises the determination of a limited number of observational units obtained from the experimental environment in such a way as to obtain the best quality of the system responses.

The importance of input data selection has already been recognized in many application domains [178]. Fault detection and isolation of industrial systems is an example which is particularly stimulating in the light of the results reported in this monograph. One of the tasks of failure protection systems is to provide reliable diagnosis of the expected system state. But to produce such a forecast, an accurate model is necessary together with its calibration, which requires parameter estimation. The preparation of experimental conditions in order to gather informative measurements can be very expensive or even impossible (e.g. for faulty system states). On the other hand, data from a real-world system may be very noisy and using all the data available may lead to significant systematic modelling errors. As a result, we are faced with the problem of how to optimise the training data in order to obtain the most precise model.