Function Approximation
This NeuronDotNet application demonstrates the use of backpropagation neural networks in function approximation problems.
Download LinksNeural networks find applications mainly in function modelling, approximation and forecasting applications. Such applications usually require to find a function which fits best to a set of known examples. The sample provided here demonstrates the use of backpropagation neural networks in such applications.
This application takes values at six points and attempts to find the function which fits best. The training set consists of X and Y co-ordinates of the points and the test inputs consist of all X values. Hidden neuron count and the number of training cycles is obtained from the user. One can extend this program to include large number of examples and can test how the network behaviour changes with number of training cycles. (Large number of cycles lead to overtraining or overfitting).
Forecasting problems are of similar kind, where the application is provided with the past data and is expected to predict future happenings. To develop such applications, one can use similar approach but the inputs in this case will also contain recent past data (This is because, in this case the function value depends mainly on past data. The network is trained to output third entry given first and second. Given second and third, the network should output fourth entry.. and so on).
Zedgraph, an open source library is used to plot the graph.
A screenshot of the application is shown below.
