神经网络训练程序BP算法

Artificial Neural Network Training Program ANN.EXE v1.10, Copyright (c) Sommnath Kundu 1994-1996 DESCRIPTION: ----------- This artificial neural network program (ann.exe) was developed in the Department of Chemical Engineering, Calcutta University, India, as a part of a research project on the application of ANN in the field of Chemical Engineering. It may be quite helpfull to those who are doing research or interested in this field. It is a general purpose ANN training program. You can use input/output data for any system to construct and train the network. To effectively simulate the system you must include all the key parameters of the system as input or output of the net. This program is menu driven and will ask you for required inputs. All the options will be saved in a configurayion file, so you can use it later as a command line parameter. Followings are some of the usefull features of the program, 1.) Any size of the network can be constructed as permitted by DOS. 2.) Different training algorithms are available, e.g., steepest descent, conjugate gradient, DFP, BFGS and Maquardt-Levenberg. 3.) A heuristic search algorithm partially based on quadratic interpolation, developed by the author, is used to find out optimum stepsize for each iteration. This algorithm increases training speed and accuracy many times. 4.) Different mode of training can be selected, e.g., pattern, moving window and block mode of training. 5.) Activation function can be selected. 6.) 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. 7.) 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-interectively 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. SAMPLE DATA: ----------- A set of sample data files are included for a model of N-number of equal volume CSTR (constant stirred tank reactor) connected in a series having concentration CA0 (of A). At time t0 some amount solution A of concentration CA is fed into 1st tank. The concentration of A in the last tank is calculated as a function of time and number of tanks present in the series. The concentrations at different time interval are used as inputs and number of tanks as outputs to train the net. The data files included with this program are as folows: CSTR_N1.DAT ==> Training data set for CSTR model CSTR_N2.DAT ==> Test data set for CSTR model W0 ==> Initial weights W1 ==> Final weights (to be calculated) TRNG.ITE ==> Iteration record file for training data TEST.ITR ==> Iteration record file for test data CSTR.DTL ==> File to save details calculation CSTR.CFG ==> Net configuration file You can train the net for this sample data by giving net configuration file (CSTR.CFG) when asked by ANN.EXE or giving it as a command-line parameter, e.g., ann cstr.cfg. All the above data files will be generated for your data file. COPYRIGHT: --------- ANN v1.10 can be freely distributed, but it should not be used or distributed for any commercial purposeS without prior permission of the author. In case of any problem or question you can contact the author in the following addresses. Somnath kundu Office: [email protected] Residence: 44/A Madhu Roy Lane, Calcutta - 700 006, India.