balance_snn.m

神经网络的工具箱, 神经网络的工具箱,

function [alpha, ym] = balance_snn(nets, bootn, testdatasets, trdatasets)
%BALANCE_SNN Compute weights for balancing networks
%
% Syntax
%
%    [alpha, ym] = balance_snn(nets, bootn, datasets)
%
% Description
%
%  BALANCE_SNN takes
%    nets	- [1 x M] matrix of net_structs trained with 
%                 costFcn = wcf_snn.
%    bootn	- number of bootstraps
%    datasets   - [1 x M] matrix of dataset_structs containing
%                 information on the data used in training 'nets'. 
%  and returns
%    alpha	- [1 x M] matrix of weights computed by balancing;
%    ym		- [N x MU x M] matrix of networks outputs;
%
% Algorithm
%
%     See: T. Heskes; Balancing between bagging and bumping.
%     Advancesin Neural Information Processing Systems 9, pages
%     466-472, Cambridge, 1997, MIT Press. 
%
% See also
%
%     TRAIN_BOOTSTRAP_SNN, TRAIN_HALFOUT_SNN,  AVERAGE_OUTPUTS_SNN, 
%     ETESTBOOT_SNN
%

%
% Syntax also
%    [alpha, ym] = balance_snn(nets, bootn, testdatasets, trdatasets)
%

if (nargin == 3)
   trdatasets = testdatasets;
end

trdata = trdatasets(1).data;
testdata = testdatasets(1).data;

M = size(nets,2);
N = size(nets(1).biases{nets(1).numLayers},1);
MU = size(testdata.T,2);

g = getg_snn(nets(1), testdata);

stdout_snn('building H voor QP-problem');
H = zeros(M,M);
ym = zeros(N, MU, M);
for m = 1:M
    ym(:,:,m) = simff_snn(nets(1,m), testdata);
end
for m1 = 1:M 
    for m2 = m1:M
        H(m1,m2) = -0.5 * ...
           ( wef_snn(ym(:,:,m1), ym(:,:,m2), g, nets(1).costFcn.fn) + ...
             wef_snn(ym(:,:,m2), ym(:,:,m1), g, nets(1).costFcn.fn) );
    stdout_snn('.'); 
    end
end
H = H + H' - diag(diag(H));
stdout_snn('(done); ');

%A = [ones(1,M);-eye(M)];
%b = [1;zeros(M,1)];
%VLB = zeros(M,1);
%VUB = ones(M,1);
%X0 = ones(M,1)/M;
%N = 1;
Aeq = ones(1,M);
A = -eye(M);
beq = 1;
b = zeros(M,1);
LB = zeros(M,1);
UB = ones(M,1);
X0 = ones(M,1)/M;
options = optimset('display', 'off');

sigma = 0.05;
stdout_snn('bootstrapping...');
X = zeros(M, bootn);
for boot = 1:bootn
    fboot = 2*etestboot_snn(nets, trdatasets)';
%    X(:,boot) = qp(H, fboot, A, b, VLB, VUB, X0, N);
    X(:,boot) = quadprog(H, fboot, A, b, Aeq, beq, LB, UB, X0, options);
    X(:,boot) = X(:,boot)/(sum(X(:,boot),1));
    stdout_snn('.');
end
alpha = mean(X,2);
stdout_snn('(done); ');