Freeware for fast development and application of approximation type networks

Freeware for fast training, validation, and application of regression/approximation networks. Fast pruning algorithm creates a nested sequence of different size networks. Includes the multilayer perceptron (MLP), functional link network, piecewise linear network, self organizing map (SOM) and K-Means clustering. C source code for applying trained networks is provided, so users can use networks in their own applications. User-supplied txt-format training data files, containing rows of numbers, can be of any size. Example training data is also provided. Fast VB Graphics for network training error and cluster formation are included. Extensive help files are provided in the software. Numap7 is highly automated and requires very few parameter choices by the user. Advanced features include a fast MLP training algorithm (faster and better than BP and LM), input feature selection, pruning (elimination) of useless units (for MLP) and modules for PLN). Training and validation error are plotted versus network size. Utilities are provided for counting patterns, deleting columns, combining files, splitting files, calculating column mean and standard deviation, and plotting column histograms. Training data can be compressed using the discrete Karhunen-Loeve transform (KLT). This freeware version of Numap7 limits the MLP to 10 hidden units and limits the PLN to 10 clusters. Upgradable to commercial versions which lacks these limitations. The classification (decision making) version of this software, called Nuclass7, is also available. Numap7.0 was developed by the Image Processing and Neural Networks Lab of Univ. of Texas at Arlington, and by Neural Decision Lab LLC.