readme.txt

基于卡尔曼滤波器的神经网络优化

TRAINING RADIAL BASIS NEURAL NETWORKS WITH THE EXTENDED KALMAN FILTER

January 27, 2000

Dan Simon
332 Stilwell Hall
Department of Electrical Engineering
Cleveland State University
1960 East 24th Street
Cleveland, OH 44115

web: http://csaxp.csuohio.edu/~simon/
email: [e-mail]

A Radial Basis Function (RBF) network is a three-layer neural network.  See reference [1] for more references and information about RBF networks.This file describes the files that were downloaded along with this readme file.  The files include the classical Iris data that can be used to test the RBF network (and other classification algorithms), along with various m-files that can be run in the MATLAB environment. M-files are written in a very high-level language that can be easily read, almost like pseudo code. The files included with this download are as follows.

IrisX.csv - This file contains the feature vectors of the classical Iris data. The features have been normalized according to their means and standard deviations, as described in [1].  There are 150 vectors in the file, and each vector has four features.  The first 50 vectors belong to the setosa class, the next 50 belong to the versicolor class, and the last 50 belong to the virginica class. In order to read this file into an array in MATLAB and use it in the m-files, you can type the following command at MATLAB's command prompt.
>> IrisX = csvread('IrisX.csv')';

IrisY.csv - This file contains the class vectors of the Iris data. The file contains 150 vectors. The first 50 vectors are (1,0,0), indicating that the first 50 vectors in IrisX.csv belong to the setosa class. The next 50 vectors are (0,1,0), indicating that the second 50 vectors in IrisX.csv belong to the versicolor class. The last 50 vectors are (0,0,1), indicating that the last 50 vectors in IrisX.csv belong to the virginica class. In order to read this file into an array in MATLAB and use it in the m-files, you can type the following command at MATLAB's command prompt.
>> IrisY = csvread('IrisY.csv')';

RBFTest.m - Thie m-file tests an already-trained RBF network in order to determine how well the network performs.  The RBF network uses linear generator functions at the hidden layer.

RBFGrad.m - This m-file trains an RBF network using gradient descent.  The network uses linear generator functions at the hidden layer.

RBFGradIter.m - This m-file run several training and testing cycles of an RBF network using gradient descent with randomly generated training and test sets for each cycle.  This determines the "average" performance of gradient descent training.

RBFGradIterProto.m - This m-file cycles through several training and testing iterations of an RBF network trained with gradient descent.  This function iterates on the number of prototypes in the RBF network.
For each prototype count, the RBF is tested by calling the RBFGradIter function.  Some of the results in [1] were generated by typing the following command at the MATLAB prompt:
>> RBFGradIterProto([ ], [ ], 1, 15, 1, p, .01, .001)
where p = the hidden layer function parameter (2, 3, or 4) - see Eq. (5) in reference [1].

RBFKalman.m - This m-file trains an RBF network using a Kalman filter.  The network uses linear generator functions at the hidden layer.

RBFKalmanIter.m - This m-file run several training and testing cycles of an RBF network using a Kalman filter with randomly generated training and test sets for each cycle.  This determines the "average" performance of Kalman filter training.

RBFKalmanIterProto.m - This m-file cycles through several training and testing iterations of an RBF network trained with a Kalman filter.  This function iterates on the number of prototypes in the RBF network.
For each prototype count, the RBF is tested by calling the RBFKalmanIter function.  Some of the results in [1] were generated by typing the following command at the MATLAB prompt:
>> RBFKalmanIterProto([ ], [ ], 1, 15, 1, p, .001, 40, 40, 40)
where p = the hidden layer function parameter (2, 3, or 4) - see Eq. (5) in reference [1].

RBFKalmanDec.m - This m-file trains an RBF network using a decoupled Kalman filter.  The network uses linear generator functions at the hidden layer.

RBFKalmanDecIter.m - This m-file run several training and testing cycles of an RBF network using a decoupled Kalman filter with randomly generated training and test sets for each cycle.  This determines the "average" performance of decoupled Kalman filter training.

RBFKalmanDecIterProto.m - This m-file cycles through several training and testing iterations of an RBF network trained with a decoupled Kalman filter.  This function iterates on the number of prototypes in the RBF network.  For each prototype count, the RBF is tested by calling the RBFKalmanDecIter function.  Some of the results in [1] were generated by typing the following command at the MATLAB prompt:
>> RBFKalmanDecIterProto([ ], [ ], 1, 15, 1, p, .001, 40, 40, 40)
where p = the hidden layer function parameter (2, 3, or 4) - see Eq. (5) in reference [1].

In order to run these m-files,  run MATLAB and make sure that the location of the files on your hard drive is part of your MATLAB path. (For example, if you downloaded the files to the c:\rbf directory on your hard drive, type "path(path, 'c:\rbf');" at MATLAB's command prompt.) Feel free to contact Professor Simon at [e-mail] with any comments or questions.

REFERENCES

[1] D. Simon, "Training Radial Basis Neural Networks with the Extended Kalman Filter," Neurocomputing, submitted for publication, 2000.