神经网络2

随机初始化权值矩阵，继续编写学习过程：

Wh = randn(12,3)/3;

Wy = randn(2,12)/12;

C = 2000; % maximum number of training epochs

J = zeros(2,C); % Initialisation of the error function

eta = [0.003 0.1]; % Training gains

for c = 1:C

H = ones(12,N)./(1+exp(-Wh*X)); % Hidden signals (L-1 by N)

Hp = H.*(1-H); % Derivatives of hidden signal

Y = tanh(Wy*H); % Output signals (m by N)

Yp = 1 - Y.^2; % Derivatives of output signals

Ey = D - Y; % The output errors (m by K)

JJ = (sum((Ey.*Ey)'))'; % The total error after one epoch

% the performance function m by 1

delY = Ey.*Yp; % Output delta signal (m by K)

dWy = delY*H'; % Update of the output matrix

% dWy is L by m

Eh = Wy(:,1:12)'*delY % The back-propagated hidden error

% Eh is L-1 by N

delH = Eh.*Hp ; % Hidden delta signals (L-1 by N)

dWh = delH*X'; % Update of the hidden matrix

% dWh is L-1 by p

Wy = Wy+0.003*dWy; Wh = Wh+0.1*dWh;

D1(:)=Y(1,:)'; D2(:)=Y(2,:)';

surfc([X1-2 X1+2], [X2 X2], [D1 D2])

J(:,c) = JJ ;

End