This program can be used to train an artificial neural network with arbitrary input/output data. It may be helpful to those who are doing research or is interested in this field. You can use input/output data from any system to construct and train the network. To effectively simulate the system, you have to include all the key parameters of the system as input or output of the net. Though the program is menu based, there is no GUI as it was written for plain DOS and the concern was given mostly on its performance. Some of the useful features of the program are listed below:

• The size of the network that can be constructed is only limited by DOS memory.
  
  
• Different training algorithms are available, e.g., steepest descent, conjugate gradient, DFP, BFGS and Maquardt-Levenberg.
  
  
• A heuristic search algorithm, partially based on quadratic interpolation, has been used to find out optimum step size for each iteration. This algorithm increased the training speed and accuracy many times.
  
  
• Different mode of training can be selected, e.g., pattern, moving window and block mode of training.
  
  
• Activation function can be selected.
  
  
• Input and output data to the net can be scaled automatically. So you can instantly compare and verify the trained net for the test data set.
  
  
• Various intermediate results during training can be stored in the data files to facilitate statistical analysis.
  
  

There is a single executable file, ann.exe. You can run it interactively or non-interactively mode by giving configuration file name as a command-line parameter. All the data files used by the program are formatted as a text file, so you can use externally created data file, but the format should be the same.

This program can be used freely for non-commercial and educational purposes. For commercial use, please contact us first.

Download ann110.zip