Neural networks in non-linear system modelling

摘要

The principles of non-linear system modelling by neural computing are described. The theory of neural computing is not presented very deeply and attention is mostly paid to the practical aspects of neural modelling. To keep the size of this report limited only the most popular network type, multilayer perception, is presented. Some of the other network types like self-organizing maps, radial basis networks and recurrent networks are introduced briefly in the end of the report. Neural computing is a statistical method which uses measurements or other data from the system which is modelled. Knowledge of the physical phenomenon in the system is not necessary because, when training the network, the training algorithm finds appropriate connections between input and output variables. However, like in all modelling methods, a good knowledge about the system helps to get better results especially when the data used is not very good. The disadvantage of a neural network model is that it is impossible to see when looking at the model how the physical system modelled works. The inputs and outputs of the model can only be seen. The reliability of the model can be proven by testing it with new data. A neural network model can be used only in those situations which are trained to the model. A neural network model cannot be extraplated with a regression model. Both static and dynamic systems can be modelled with a neural network.