average_outputs_snn.m

神经网络的工具箱, 神经网络的工具箱,

function y_av = average_outputs_snn(ym, alpha, errf)
%AVERAGE_OUTPUTS_SNN  Weighted output average.
%
%  Syntax
%
%   y_av = average_outputs_snn(ym, alpha, errf) 
%
%  Description
%
%   AVERAGE_OUTPUTS_SNN takes
%     ym	- [NL x MU x M] matrix of networks outputs. 
%     alpha 	- [1 x M] matrix of weighting factors.
%     errf 	- string OR [N x 1] cell array of strings of errorfunction(s).
%   and returns
%     y_av      - [NL x MU] matrix of weighted outputs.
%
%   (NL = #outputs; MU = #patterns; M = #networks in ensemble)
%
%   Supported output errorfunctions are:
%     - SE_SNN              squared error;
%     - RELERR_SNN          relative error;
%     - CROSSENTROPY_SNN    cross entropy;
%     - CROSSLOGISTIC_SNN   cross logisitc;
%     - LOGLIKELIHOOD_SNN   log likelihood;
%
% See also
%
%     BALANCE_SNN, SIMFF_AVR_SNN
%

if (nargin ~= 3)
   error('Must have 3 input arguments.');
end

[N, MU, M] = size(ym);
delta = 1e-6;

alpha = alpha/(sum(alpha,1));

if isstr(errf)
   switch (errf)
     case {'se_snn'},
       y_av = reshape(reshape(ym, N*MU, M) * alpha, N, MU); 
     case {'qe_snn'},
       y_av = reshape(reshape(ym, N*MU, M) * alpha, N, MU); 
     case {'relerr_snn'},
       y_av = reshape(...
                   ((reshape(ym, N*MU, M).^2) * alpha)./ ...
                   (reshape(ym, N*MU, M) * alpha), ...
	           N, MU);
     case {'crossentropy_snn'},
       y_av = reshape(...
                 tanh(atanh(max(min(reshape(ym, N*MU, M), 1-delta),... 
		      -1+delta))...
                 * alpha), ...
		 N, MU);
     case {'crosslogistic_snn'},
       ym = reshape(max(min(ym, 1-delta), delta), N*MU, M);
       y_av = reshape(...
                 1./(1 + exp(log((1-ym)./ym) * alpha)), ...
                 N, MU);
     case {'loglikelihood_snn'},
       y_av = reshape(1./((1./reshape(ym, N*MU, M)) * alpha), N, MU);
     otherwise,
       error(sprintf('No support for output error function %s', errf));
   end
elseif iscell(errf)
       y_av = zeros(N, MU);
       for i = 1:N
	   switch (errf{i,1})
             case {'se_snn'},
               y_av(i, :) = (reshape(ym(i,:,:), MU, M) * alpha)'; 
             case {'qe_snn'},
               y_av(i, :) = (reshape(ym(i,:,:), MU, M) * alpha)'; 
             case {'relerr_snn'},
               y_av(i, :) = ... 
                   ( ((reshape(ym(i,:,:), MU, M).^2) * alpha)./ ...
                   (reshape(ym(i,:,:), MU, M) * alpha) )';
             case {'crossentropy_snn'},
               y_av(i, :) = ...
                 (tanh(atanh(max(min(reshape(ym(i,:,:), MU, M), 1-delta),...
		      -1+delta))...
                 * alpha))';
             case {'crosslogistic_snn'},
               tmp = reshape(max(min(ym(i,:,:), 1-delta), delta), MU, M);
               y_av(i, :) = (1./(1 + exp(log((1-tmp)./tmp) * alpha)))';
             case {'loglikelihood_snn'},
               y_av(i, :) = (1./((1./reshape(ym(i,:,:), MU, M)) * alpha))';
             otherwise,
               error(sprintf(...
	           'No support for output errorfunction %s', errf{i,1}));
	   end
       end
else
   error('Incorrect format for output errorfunction');
end