(code posted)
(make code work with general mean and covariance matrix)
 
Line 21: Line 21:
 
% e --- Bayes Error
 
% e --- Bayes Error
  
x1 = randn(n,N);
+
 
 +
 
 +
temp = randn(n,N);
 +
[A, B] = eig(Sigma1);
 +
x1 = A*sqrt(B)*temp + repmat(M1, 1, N);
 
y1 = discFunc(x1, P1, M1, Sigma1, P2, M2, Sigma2, 1);  
 
y1 = discFunc(x1, P1, M1, Sigma1, P2, M2, Sigma2, 1);  
 
e1 = y1/N;
 
e1 = y1/N;
Line 27: Line 31:
  
 
temp = randn(n,N);
 
temp = randn(n,N);
x2 = sqrt(Sigma2)*temp + repmat(M2, 1, N);
+
[A, B] = eig(Sigma2);
 +
x2 = A*sqrt(B)*temp + repmat(M2, 1, N);
 
y2 = discFunc(x2, P1, M1, Sigma1, P2, M2, Sigma2, 2);
 
y2 = discFunc(x2, P1, M1, Sigma1, P2, M2, Sigma2, 2);
 
e2 = y2/N;
 
e2 = y2/N;
Line 36: Line 41:
  
 
end
 
end
 +
 +
 
</source>
 
</source>
  

Latest revision as of 17:50, 11 February 2016


Function to compute Bayes error (based on Monte-Carlo method)

function [ e ] = BayesErrorMonteCarlo( n, P1, M1, Sigma1, P2, M2, Sigma2, N )
%Compute Bayes Error using Monte-Carlo method
%J. Zhou
 
% ....Input Parameters
% n --- data dimension
% P1 --- prior probability of class 1
% M1 --- mean of class 1 data
% Sigma1 --- covariance of class 1 data
% p2 --- prior probability of class 2
% M2 --- mean of class 2 data
% Sigma2 --- covariance of class 2 data
% N --- number of samples using in Monte-Carlo
 
%....Output Parameters
% e --- Bayes Error
 
 
 
temp = randn(n,N);
[A, B] = eig(Sigma1);
x1 = A*sqrt(B)*temp + repmat(M1, 1, N);
y1 = discFunc(x1, P1, M1, Sigma1, P2, M2, Sigma2, 1); 
e1 = y1/N;
 
 
temp = randn(n,N);
[A, B] = eig(Sigma2);
x2 = A*sqrt(B)*temp + repmat(M2, 1, N);
y2 = discFunc(x2, P1, M1, Sigma1, P2, M2, Sigma2, 2);
e2 = y2/N;
 
% Bayes Error e
e = P1*e1+P2*e2;
 
 
end

Function #1 used in previous code

function [ z ] = discFunc(x,P1,M1,Sigma1,P2,M2,Sigma2, CFlag)
%compute 3.119 and return number of samples that are misclassified
%J. Zhou
 
[N,M] = size(x); %M samples
 
invSigma1 = inv(Sigma1);
invSigma2 = inv(Sigma2);
temp = 0.5*log(det(Sigma1)/det(Sigma2)) - log(P1/P2);
 
y = zeros(1,M);
 
for i = 1:M          
 
        if  0.5*(x(:,i)-M1)'*invSigma1*(x(:,i)-M1) - ...
                0.5*(x(:,i)-M2)'*invSigma2*(x(:,i)-M2) + temp > 0   
            if CFlag == 1
                y(i) = 1;
            else
                y(i) = 0;
            end
        else
            if CFlag == 1
                y(i) = 0;
            else
                y(i) = 1;
            end
        end
 
end
 
z = sum(y);
 
end

Test Function

%% ECE662 
% This code tests BayesErrorMonteCarlo.m using a data set with known
% Bayes Error
% J. Zhou
 
close all;
clear all;
 
% test data using Data I-^ ; Fukunaga page 45
n = 8;   %data dimension
 
P1 = 0.5; % prior probability for class 1
M1 = zeros(n, 1);  %mean of class 1 data
Sigma1 = eye(n,n);  %covariance of class 1 data
 
P2 = 0.5; % prior probability for class 2
M2 = [3.86, 3.10, 0.84, 0.84, 1.64, 1.08, 0.26, 0.01]';
Sigma2 = diag([8.41, 12.06, 0.12, 0.22, 1.49, 1.77, 0.35, 2.73]);
 
 
%Number of samples for Monte-Carlo, start with 1 million, run a few times
%to see how much computed Bayes error vary, if it varies quite bit, then
%increase the number of samples, do this until Bayes error become
%relatively stable. 
N = 1000000;  
 
e = BayesErrorMonteCarlo(n, P1, M1, Sigma1, P2, M2, Sigma2, N );
 
% Bayes Error in this case should 1.46%
sprintf('Bayes Error in this case should be %.3f', 0.018)
 
sprintf('Bayes Error computed is %.3f', e)

Alumni Liaison

Basic linear algebra uncovers and clarifies very important geometry and algebra.

Dr. Paul Garrett