nmatreg.m

来自「这是一个基于线性矩阵的神经网络控制程序」· M 代码 · 共 80 行

%
%      Linear neural network with matrix inputs.
%
% In this model we take sums of u'Xv and iteratively optimize parameters of
% the sums. For paper(s) see pd.hacker.lt
%
% Model created and implemented by Povilas Daniu餴s, [e-mail],
% 2008
%
% usage nMatReg(X,y,M,N,lambda)
%
% where X is an input matrix data structure (i.e. X(i).mat would be a i-th
% element
% y     - is corresponding target values
% M     - iterations
% N     - number of regressors
% labda - regularization constant

function [u,v]=nMatReg(X,y,M,N,lambda)


[rows,cols] = size(X(1).mat);

for i=1:N
    u(:,i) = rand(rows,1);
    v(:,i) = rand(cols,1);
end    


for i=1:M % iterations
    
    for k=1:N % order
              
        % find u(k)
        A = zeros(rows,rows); 
        z = zeros(rows,cols);             
        for j=1:length(y)
            A = A + X(j).mat * v(:,k) *  v(:,k)' *  X(j).mat';  
            
                akx = 0;
                for m=1:N
                    if (m ~= k)
                        akx = akx + v(:,m)'*X(j).mat'*u(:,m);             
                    end 
                end          
            
            z = z + (y(j) - akx) * X(j).mat;
        end                                           
        u(:,k) = inv(A + lambda*eye(rows)) * z * v(:,k);
      
            
        % find v(k)
        A = zeros(cols,cols); 
        z = zeros(cols,rows);             
        for j=1:length(y)
            A = A + X(j).mat' * u(:,k) *  u(:,k)' *  X(j).mat;  
            
                bkx = 0;
                for m=1:N
                    if (m ~= k)
                        bkx = bkx + u(:,m)'*X(j).mat*v(:,m);             
                    end 
                end                      
            z = z + (y(j) - bkx) * X(j).mat';
        end                                           
        v(:,k) = inv(A + lambda*eye(cols)) * z * u(:,k);
        
    % test SSE
    
    sse = 0;
    for jj=1:length(y)
        out = 0;
        for kk=1:N
            out = out + u(:,kk)'*X(jj).mat*v(:,kk);
        end
        sse = sse + (y(jj) - out).^2;
    end    
    fprintf('Iteration %d, regressor %d, SSE = %f \n',i,k,sse);         
    end       
end