www.gusucode.com > nncontrol 工具箱 matlab 源码程序 > nncontrol/private/nnmodrefhelp_train_contr.m
% TRAINING THE CONTROLLER % % Before training the controller, a neural network plant model must first % be correctly identified. If you have not previously identified the plant, % then click the Plant Identification button, which will open an identification % window. % % The controller training algorithm needs the following parameters: % % 1) Size of the Hidden layer: Define how many neurons will be in the hidden % layer of the controller. % 2) Reference Model: A simulink file, with inport and outport blocks, used to % generate a reference response to train the controller. % 3) No. Delayed Reference Inputs: defines how many delays in the reference % will be used to feed the controller. % 4) No. Delayed Controller Outputs: defines how many delays in the controller % output will be used to feed the controller. % 5) No. Delayed Plant Outputs: defines how many delays in the plant output % will be used to feed the controller. % 6) Maximum/Minimum Reference Values: Defines the bounds on the random % input to the reference model. % 7) Maximum/Minimum Interval Values: Defines the bounds on the interval % over which the random reference will remain constant. % 8) Controller Training Samples: Defines the number of random values to % be generated to feed the reference model and therefore to be used % in training the controller. % 9) Controller Training Epochs: Defines how many epochs per segment will % be used during training. One segment of data is presented to the network, % and then the specified number of epochs of training are performed. % The next segment is then presented, and the process is repeated. This % continues until all segments have been presented. % 10) Controller Training Segments: Defines how many segments the training data % is divided into. % 11) Use Cumulative Training: If selected, the initial training is done with % one segment of data. Future training is performed by adding segments % to the previous training data, until the entire training data set is % used in the final stage. Use this option carefully due to increased training % time. % 12) Use Current Weights: If selected, the current controller weights % are used as the initial weights for controller training. % Otherwise, random initial weights are generated. % If the controller network structure is modified, this option % will be overridden, and random weights will be used. % % The Generate Training Data button generates training data based on the % reference model file. You can also Import training data. Once the training % data is entered, you can perform one of the following actions: % % 1) Train Controller: Trains the neural network controller using % the available data. The previous weights are used as initial weights, % if that option is selected. % 2) Apply: The weights are saved in the Neural Network Controller block. % You can simulate the system while this window remains open. % 3) OK: The weights are saved in the Neural Network Controller block, and % the window is closed. % 4) Cancel: The window is closed, and no values are saved. % 5) Plant Identification: Opens a Plant Identification window. % You should identify the plant before performing controller % training. You may also want to re-identify the plant if the % controller training is not satisfactory. An accurate plant % model is needed for accurate controller training. % % During the training process, progress report messages are shown in the % feedback line. % Copyright 1992-2013 The MathWorks, Inc.