gusucode.com > 用matlab仿真0到9十个数字的语音识别源码程序 > Speech-Recogenition/em_converged.m
function [converged, decrease] = em_converged(loglik, previous_loglik, threshold, check_increased) % EM_CONVERGED Has EM converged? % [converged, decrease] = em_converged(loglik, previous_loglik, threshold) % % We have converged if the slope of the log-likelihood function falls below 'threshold', % i.e., |f(t) - f(t-1)| / avg < threshold, % where avg = (|f(t)| + |f(t-1)|)/2 and f(t) is log lik at iteration t. % 'threshold' defaults to 1e-4. % % This stopping criterion is from Numerical Recipes in C p423 % % If we are doing MAP estimation (using priors), the likelihood can decrase, % even though the mode of the posterior is increasing. if nargin < 3, threshold = 1e-4; end if nargin < 4, check_increased = 1; end converged = 0; decrease = 0; if check_increased if loglik - previous_loglik < -1e-3 % allow for a little imprecision fprintf(1, '******likelihood decreased from %6.4f to %6.4f!\n', previous_loglik, loglik); decrease = 1; converged = 0; return; end end delta_loglik = abs(loglik - previous_loglik); avg_loglik = (abs(loglik) + abs(previous_loglik) + eps)/2; if (delta_loglik / avg_loglik) < threshold, converged = 1; end