function [LL, init_state_prob, transmat, mu, Sigma, mixmat, mu_history] = ...    learn_mhmm(data, init_state_prob, transmat, mu, Sigma, mixmat, ...    max_iter, thresh, verbose, cov_type, static)% LEARN_MHMM Compute the ML parameters of an HMM with mixtures of Gaussians output using EM.% % [LL, PRIOR, TRANSMAT, MU, SIGMA, MIXMAT, MU_HISTORY] = LEARN_MHMM(DATA, PRIOR0, TRANSMAT0,% MU0, SIGMA0, MIXMAT0) computes the ML estimates of the following parameters,% where, for each time t, Q(t) is the hidden state, M(t) is the mixture component, and Y(t) is% the observation.%   prior(i) = Pr(Q(1) = i), %   transmat(i,j) = Pr(Q(t+1)=j | Q(t)=i)%   mixmat(j,k) = Pr(M(t)=k | Q(t)=j) %   mu(:,j,k) = E[Y(t) | Q(t)=j, M(t)=k ]%   Sigma(:,:,j,k) = Cov[Y(t) | Q(t)=j, M(t)=k]% PRIOR0 is the initial estimate of PRIOR, etc.% To learn an HMM with a single Gaussian output, just set mixmat = ones(Q,1).%% DATA(:,t,l) is the observation vector at time t for sequence l. If the sequences are of% different lengths, you can pass in a cell array, so DATA{l} is an O*T matrix.%% LL is the "learning curve": a vector of the log lik. values at each iteration.% LL might go positive, since prob. densities can exceed 1, although this probably% indicates that something has gone wrong e.g., a variance has collapsed to 0.% MU_HISTORY(:,:,:,r) is MU at iteration r.%% There are several optional arguments, which should be passed in the following order%  LEARN_MHMM(DATA, INIT_STATE_PROB, TRANSMAT, MU, SIGMA, MIXMAT, ...%    MAX_ITER, THRESH, VERBOSE, COV_TYPE, STATIC)% These have the following meanings%% max_iter = max. num EM steps to take (default 10)% thresh = threshold for stopping EM (default 1e-4)% verbose = 0 to suppress the display of the log lik at each iteration (Default 1).% cov_type = 'full', 'diag' or 'spherical' (default 'full')% static = 1 if we don't want to re-estimate prior and transmat, i.e., no dynamics (default = 0)%% See "A tutorial on Hidden Markov Models and selected applications in speech recognition",% L. Rabiner, 1989, Proc. IEEE 77(2):257--286.%learn_mhmm(data, init_state_prob, transmat, mu, Sigma, mixmat, max_iter, thresh, verbose, cov_type, static)if nargin<7, max_iter = 10; endif nargin<8, thresh = 1e-4; endif nargin<9, verbose = 1; endif nargin<10, cov_type = 'full'; endif nargin<11, static = 0; endif ~iscell(data)  data = num2cell(data, [1 2]); % each elt of the 3rd dim gets its own cellendnumex = length(data);  previous_loglik = -inf;loglik = 0;converged = 0;iter = 1;LL = [];O = size(data{1},1);Q = length(init_state_prob);M = size(mixmat,2);mu_history = zeros(O,Q,M,max_iter);while (iter <= max_iter) & ~converged  % E step  [loglik, exp_num_trans, exp_num_visits1, postmix, m, ip, op] = ...      ess_mhmm(init_state_prob, transmat, mixmat, mu, Sigma, data);  if verbose, fprintf('iteration %d, loglik = %f\n', iter, loglik); end  % M step  if ~static    init_state_prob = normalise(exp_num_visits1);    transmat = mk_stochastic(exp_num_trans);  end  if M == 1    mixmat = ones(Q,1);  else    mixmat = mk_stochastic(postmix);  end  for j=1:Q    for k=1:M      mu(:,j,k) = m(:,j,k) / postmix(j,k);      if cov_type(1) == 's'	s2 = (1/O)*( (ip(j,k)/postmix(j,k)) - mu(:,j,k)'*mu(:,j,k) );	Sigma(:,:,j,k) = s2 * eye(O);      else	SS = op(:,:,j,k)/postmix(j,k) - mu(:,j,k)*mu(:,j,k)';	if cov_type(1)=='d'	  SS = diag(diag(SS));	end	Sigma(:,:,j,k) = SS;      end    end  end    mu_history(:,:,:,iter) = mu;  ll_history(iter) = loglik;  converged = em_converged(loglik, previous_loglik, thresh);  previous_loglik = loglik;  iter =  iter + 1;  LL = [LL loglik];end%%%%%%%%%function [loglik, exp_num_trans, exp_num_visits1, postmix, m, ip, op] = ...    ess_mhmm(prior, transmat, mixmat, mu, Sigma, data)% ESS_MHMM Compute the Expected Sufficient Statistics for a MOG Hidden Markov Model.%% Outputs:% exp_num_trans(i,j)   = sum_l sum_{t=2}^T Pr(Q(t-1) = i, Q(t) = j| Obs(l))% exp_num_visits1(i)   = sum_l Pr(Q(1)=i | Obs(l))% postmix(i,k) = sum_l sum_t w(i,k,t) where w(i,k,t) = Pr(Q(t)=i, M(t)=k | Obs(l))  (posterior mixing weights)% m(:,i,k)   = sum_l sum_t w(i,k,l) * Obs(:,t,l)% ip(i,k) = sum_l sum_t w(i,k,l) * Obs(:,t,l)' * Obs(:,t,l)% op(:,:,i,k) = sum_l sum_t w(i,k,l) * Obs(:,t,l) * Obs(:,t,l)'%% where Obs(l) = Obs(:,:,l) = O_1 .. O_T for sequence lverbose = 0;%[O T numex] = size(data);numex = length(data);O = size(data{1},1);Q = length(prior);M = size(mixmat,2);exp_num_trans = zeros(Q,Q);exp_num_visits1 = zeros(Q,1);postmix = zeros(Q,M);m = zeros(O,Q,M);op = zeros(O,O,Q,M);ip = zeros(Q,M);loglik = 0;if verbose, fprintf(1, 'forwards-backwards example # '); endfor ex=1:numex  if verbose, fprintf(1, '%d ', ex); end  %obs = data(:,:,ex);  obs = data{ex};  T = size(obs,2);  [B, B2] = mk_mhmm_obs_lik(obs, mu, Sigma, mixmat);  [gamma, xit, current_loglik, gamma2] = forwards_backwards_mix(prior, transmat, B, B2, mixmat);  loglik = loglik +  current_loglik;   if verbose, fprintf(1, 'll at ex %d = %f\n', ex, loglik); end  exp_num_trans = exp_num_trans + sum(xit,3);  exp_num_visits1 = exp_num_visits1 + gamma(:,1);  postmix = postmix + sum(gamma2,3);  for j=1:Q    for k=1:M      w = reshape(gamma2(j,k,:), [1 T]); % w(t) = Pr(Q(t)=j, M(t)=k | obs)      wobs = obs .* repmat(w, [O 1]); % wobs(:,t) = w(t) * obs(:,t)      m(:,j,k) = m(:,j,k) + sum(wobs, 2); % m(:) = sum_t w(t) obs(:,t)      op(:,:,j,k) = op(:,:,j,k) + wobs * obs'; % op(:,:) = sum_t w(t) * obs(:,t) * obs(:,t)'      ip(j,k) = ip(j,k) + sum(sum(wobs .* obs, 2)); % ip = sum_t w(t) * obs(:,t)' * obs(:,t)    end  endendif verbose, fprintf(1, '\n'); end