The training sample is provided as a training algorithm with teachers. The final acquisition of the network model requires repeated iterative calculation based on a certain amount of experimental data to make the network output approach the target value. One rule of thumb is that the training sample should be 510 times the total number of connection rights. A large amount of sample data is stored in data files, including input vector and target vector data files. Data variables can be loaded from disk using MATLAB with the load*1dat command.
Network Initialization In the BP network, as long as the input vector p, the target vector t, the number of neurons in each layer, and the transfer function of each layer of neurons are known, the BPIT can be initialized using the function INITFF.
Network Training The network training function provided by the MATLAB Neural Network Toolbox can be used directly to train a BP neural network in a loop, and finally achieve a function that allows the target error. The commonly used training function is TRAINBP, which is different from the learning rules used. The function trainbp uses the standard BP learning rules to train the feedforward network, so that the network can complete function approximation, vector classification and pattern recognition. Trainbpx adopts the momentum method and the learning rate adaptive adjustment strategy, which improves the learning speed and increases the training forward. The network has a fast algorithm and a short learning time, but it requires more memory, especially for complex problems. The following program shows the trainbpx calling method: training process display frequency, maximum training steps, error index, learning speed, learning rate increase ratio, learning rate reduction ratio, momentum factor, maximum error ratio ratio2, training parameter casting, etc.: based on MATLAB's casting defect analysis and control has been trained to obtain a new weight matrix W1, W2. (End)