Image Processing and Neural Networks Lab, The University of Texas at Arlington. http://www.uta.edu/faculty/manry/

This program trains a multi-layer perceptron with one hidden layer. The training algorithm used is Adaptive HWO-MOLF. See: Rohit Rawat, Jignesh Patel and Michael Manry, "Minimizing Validation Error With Respect to Network Size and Number of Training Epochs," the 2013 International Joint Conference on Neural Networks.

The training data has N inputs x and M outputs t. Patterns are arranged in rows with a tab or space separating the elements.

x_1 x_2 ... x_N t_1 t_2 ... t_M

Example data file with N = 4, M = 2 0.8147 0.0975 0.1576 0.1419 0.6557 0.7577 0.9058 0.2785 0.9706 0.4218 0.0357 0.7431 0.1270 0.5469 0.9572 0.9157 0.8491 0.3922 0.9134 0.9575 0.4854 0.7922 0.9340 0.6555 0.6324 0.9649 0.8003 0.9595 0.6787 0.1712 The first four columns are the inputs and the last two columns are the outputs.

The training data has N inputs x and M classes ic. Patterns are arranged in rows with a tab or space separating the elements. The class numbers ic must start from 1, going all the way to M.

x_1 x_2 ... x_N ic

Example data file with N = 4, Nc = 2 0.6020 0.0838 0.9961 0.7749 2.0000 0.2630 0.2290 0.0782 0.8173 1.0000 0.6541 0.9133 0.4427 0.8687 2.0000 0.6892 0.1524 0.1067 0.0844 2.0000 0.7482 0.8258 0.9619 0.3998 1.0000 0.4505 0.5383 0.0046 0.2599 1.0000 The first four columns are the inputs and the last column has the class numbers.

This program uses validation data for network sizing. If you have already split your data into training and validation files, supply both files to the program. If you did not, the program can automatically split the training file to generate validation data.

N: inputs, M: outputs, Nh: the number of hidden units to begin with, Nit: the maximum number of training iterations/epochs. During training, the number of hidden units and iterations is optimized to minimize the validation error.

Run the program: "run_training.m" Configure all the parameters in the GUI and press the "Train" button. The program outputs can be seen in the MATLAB console. The network weights are stored in the file weights.txt.

The console version of the program is provided in the hwo_molf_pruning folder. please run the program "mlp_TRAIN.m", or "mlp_TRAIN_CLASS.m" to train a regression or a classification model. Running these programs without any arguments prompts the user for input on the command line. These can also be invoked as functions and their prototypes are given below:

[E_t_best E_v_best Nh_best Nit_best Wi_best Wo_best lambda] = mlp_TRAIN(training_file, N, M, Nh, Nit, validation_file)

[E_t_best E_v_best Nh_best Nit_best Wi_best Wo_best lambda] = mlp_TRAIN_CLASS(training_file, N, M, Nh, Nit, validation_file)

The best network size and training iterations are returned as Nh_best and Nit_best. The training and validation error for the best network is returned as E_t_best and E_v_best.

Please run "mlp_PROCESSING.m" or "mlp_PROCESSING_CLASS.m" to obtain processing results on new data. If the new data has the correct outputs available, a testing error is calculated and displayed. The GUI for the testing program is "run_testing.m".

README version 1 Rohit Rawat 08/24/2015