| dks | Date: Friday, 08 March 13, 5:00 PM | Message # 1 |
|
New Lit
Group: Friends
Messages: 4
Status: Offline
| % weight_vector = perceptron(InputSamples,TrainingRate)
% the perceptron training algorithm pg 47 Elements of ANN
% Given Input, a set of training input vectors of the form
% [X0,X1,...Xn-1,Class]
% where X0 is a constant 1 and Class is -1 for Class0 or +1 for Class1
% and a TrainingRate = (0,1]
% returns weight a vector of the form
% [W0,W1,...Wm] : m = n-1
% such that the weight vector dichotomizes Input. If the training samples
% are not linearly seperable it does not terminate. In other words
% determines (approximates) a weight vector such that
% W0 + W1*X1 + ,..., + Wm*Xn-1 = 0
function w = perceptrona(Input,n)
Ip = Input(:,1:end-1); % Ip: the sample x's
c = Input(:,end); % c: vector of corresponding classes
[r,cols] = size(Ip); % cols: number of columns in Ip
w{1} = rand(1,cols)'; % create random weight vector of size cols
k = 2; % using 1-based indexing set k to 2
% while there exist input vectors in Ip that are missclassified
% by w{k-1} pick a missclassifed Ij in Ip and adjust w{k-1} by delta w
% to get w{k}
i = misclassified(Ip,w{k-1},c,r);
while i
xk = c(i)*Ip(i,:);
w{k} = w{k-1} + n*xk';
k = k + 1;
i = misclassified(Ip,w{k-1},c,r);
end
w = w{k-1}'; % return the column vector
% result = misclassified(I,w,c,n)
% returns the index of the first misclassified input vector
% in I if no input vectors in I are misclassifed returns 0
function result = misclassified(I,w,c,n)
result = 0;
ij = I * w;
for i = 1:n
if (ij(i) <= 0 && c(i) ~= -1) || (ij(i) > 0 && c(i) ~= 1)
result = i;
return;
end
end
|
| |
| |
