weird performance of pattern classifier using matlab

[mamech]

hello everyone

I am newcomer to the world of AI, and I tried to make a matlab code where the user enter 4 points on the graph, and the program draw a line that separate the last point (which is supposed to be always class 1) from the other 3 points (supposed always to be class 0).
The main problem that I found bad performance from the program. In the beginning, I tested the program in learning AND gate, and it works very fine and it could get the right answer may be only in 10 or 100 iterations, but once I made it take input from the user, I found that the program sometimes classifies right and sometimes wrong even with 10000 iterations!!!

Note: I select the 4 points on the graph always in form of linearly separable formation, so we can exclude the misclassification because of this mistake

close all
clear
clc
Xa=[];
Ya=[];
Xb=[];
Yb=[];
input=[];
num_ip = 4;
bias = -1;
learning_rate = 0.1;
weights = rand(3,1);
z=0:0.1:10
iterations = 1000


while(1)
    if(length(Xa)==4)
            break;
        end
    axis([0 10 0 10]); % open an axis from 0__>10 in each axis
    grid on
    [x,y]=ginput(1); % to get input from user .the programm Continue to get input from user
    % until you press enter .
    plot(x,y,'ro'); % plot the point 
    hold on
    
    if(isempty(x)) % if you press enter the programm break
       break;
    end   
        Xa(end+1) = x; % to save the x data in a vector
        Ya(end+1) = y; % to save the y data in a vector
        
        
end


tempx=[Xa  ]'
tempy= [Ya]';
input=[tempx tempy];
       
desired_out = [0;0;0;1];

for i = 1:iterations
     out = zeros(4,1);
     for j = 1:num_ip
          y = weights(1,1)*input(j,1)+ weights(2,1)*input(j,2)+weights(3,1)*bias;
          if(sign(y)==0 || sign(y)==1 )
          out(j) = 1 ;
          end
          if(sign(y)==-1)
          out(j) = 0 ;
          end
          delta = desired_out(j)-out(j);
         
         
          weights(1,1) = weights(1,1)+learning_rate*input(j,1)*delta;
          weights(2,1) = weights(2,1)+learning_rate*input(j,2)*delta;
          weights(3,1) = weights(3,1)+learning_rate*bias*delta;
         
          

hold on 



     end
       
     
end


classifier= -(weights(2,1)/weights(1,1))*z - bias*(weights(3,1)/weights(1,1));
plot(z,classifier)

 

[andre_luis]

You must try other kind of classifiers in order to take a reference to compare.
One of the most promising ones is the SVM, avaliable on link bellow in many languages:

https://www.csie.ntu.edu.tw/~cjlin/libsvm/


+++

 

[mamech]

thank you for your help, I discovered where is the error. It was in the equation of the line that represent the classifier barrier. I defined the equation wrongly, so it gave me a wrong impression that the whole process is wrong.