Analysis of networks that have learnt control problems

Shows how an artificial neural network which has been trained to approximate a specific mapping can be reduced to a single equation of the inputs. To do this the authors have considered the output functions of the networks units in terms of their Taylor series expansion, and also introduced an alternative output function to make the process easier. Although this numerical technique produces an equation for the mapping, it is hard to see how this equation can give an insight into the analytical relationship between the input and output spaces. One place that the technique may prove fruitful is in the hardware implementation of networks. So far this field has been hampered by the problem of large numbers of connections, but this method reduces that highly parallel architecture into a simple equation that may be able to be implemented in hardware