Daniel@0: function [A, C, Q, R, initx, initV, LL] = ...
Daniel@0:     learn_kalman(data, A, C, Q, R, initx, initV, max_iter, diagQ, diagR, ARmode, constr_fun, varargin)
Daniel@0: % LEARN_KALMAN Find the ML parameters of a stochastic Linear Dynamical System using EM.
Daniel@0: %
Daniel@0: % [A, C, Q, R, INITX, INITV, LL] = LEARN_KALMAN(DATA, A0, C0, Q0, R0, INITX0, INITV0) fits
Daniel@0: % the parameters which are defined as follows
Daniel@0: %   x(t+1) = A*x(t) + w(t),  w ~ N(0, Q),  x(0) ~ N(init_x, init_V)
Daniel@0: %   y(t)   = C*x(t) + v(t),  v ~ N(0, R)
Daniel@0: % A0 is the initial value, A is the final value, etc.
Daniel@0: % DATA(:,t,l) is the observation vector at time t for sequence l. If the sequences are of
Daniel@0: % different lengths, you can pass in a cell array, so DATA{l} is an O*T matrix.
Daniel@0: % LL is the "learning curve": a vector of the log lik. values at each iteration.
Daniel@0: % LL might go positive, since prob. densities can exceed 1, although this probably
Daniel@0: % indicates that something has gone wrong e.g., a variance has collapsed to 0.
Daniel@0: %
Daniel@0: % There are several optional arguments, that should be passed in the following order.
Daniel@0: % LEARN_KALMAN(DATA, A0, C0, Q0, R0, INITX0, INITV0, MAX_ITER, DIAGQ, DIAGR, ARmode)
Daniel@0: % MAX_ITER specifies the maximum number of EM iterations (default 10).
Daniel@0: % DIAGQ=1 specifies that the Q matrix should be diagonal. (Default 0).
Daniel@0: % DIAGR=1 specifies that the R matrix should also be diagonal. (Default 0).
Daniel@0: % ARMODE=1 specifies that C=I, R=0. i.e., a Gauss-Markov process. (Default 0).
Daniel@0: % This problem has a global MLE. Hence the initial parameter values are not important.
Daniel@0: % 
Daniel@0: % LEARN_KALMAN(DATA, A0, C0, Q0, R0, INITX0, INITV0, MAX_ITER, DIAGQ, DIAGR, F, P1, P2, ...)
Daniel@0: % calls [A,C,Q,R,initx,initV] = f(A,C,Q,R,initx,initV,P1,P2,...) after every M step. f can be
Daniel@0: % used to enforce any constraints on the params. 
Daniel@0: %
Daniel@0: % For details, see
Daniel@0: % - Ghahramani and Hinton, "Parameter Estimation for LDS", U. Toronto tech. report, 1996
Daniel@0: % - Digalakis, Rohlicek and Ostendorf, "ML Estimation of a stochastic linear system with the EM
Daniel@0: %      algorithm and its application to speech recognition",
Daniel@0: %       IEEE Trans. Speech and Audio Proc., 1(4):431--442, 1993.
Daniel@0: 
Daniel@0: 
Daniel@0: %    learn_kalman(data, A, C, Q, R, initx, initV, max_iter, diagQ, diagR, ARmode, constr_fun, varargin)
Daniel@0: if nargin < 8, max_iter = 10; end
Daniel@0: if nargin < 9, diagQ = 0; end
Daniel@0: if nargin < 10, diagR = 0; end
Daniel@0: if nargin < 11, ARmode = 0; end
Daniel@0: if nargin < 12, constr_fun = []; end
Daniel@0: verbose = 1;
Daniel@0: thresh = 1e-4;
Daniel@0: 
Daniel@0: 
Daniel@0: if ~iscell(data)
Daniel@0:   N = size(data, 3);
Daniel@0:   data = num2cell(data, [1 2]); % each elt of the 3rd dim gets its own cell
Daniel@0: else
Daniel@0:   N = length(data);
Daniel@0: end
Daniel@0: 
Daniel@0: N = length(data);
Daniel@0: ss = size(A, 1);
Daniel@0: os = size(C,1);
Daniel@0: 
Daniel@0: alpha = zeros(os, os);
Daniel@0: Tsum = 0;
Daniel@0: for ex = 1:N
Daniel@0:   %y = data(:,:,ex);
Daniel@0:   y = data{ex};
Daniel@0:   T = length(y);
Daniel@0:   Tsum = Tsum + T;
Daniel@0:   alpha_temp = zeros(os, os);
Daniel@0:   for t=1:T
Daniel@0:     alpha_temp = alpha_temp + y(:,t)*y(:,t)';
Daniel@0:   end
Daniel@0:   alpha = alpha + alpha_temp;
Daniel@0: end
Daniel@0: 
Daniel@0: previous_loglik = -inf;
Daniel@0: loglik = 0;
Daniel@0: converged = 0;
Daniel@0: num_iter = 1;
Daniel@0: LL = [];
Daniel@0: 
Daniel@0: % Convert to inline function as needed.
Daniel@0: if ~isempty(constr_fun)
Daniel@0:   constr_fun = fcnchk(constr_fun,length(varargin));
Daniel@0: end
Daniel@0: 
Daniel@0: 
Daniel@0: while ~converged & (num_iter <= max_iter) 
Daniel@0: 
Daniel@0:   %%% E step
Daniel@0:   
Daniel@0:   delta = zeros(os, ss);
Daniel@0:   gamma = zeros(ss, ss);
Daniel@0:   gamma1 = zeros(ss, ss);
Daniel@0:   gamma2 = zeros(ss, ss);
Daniel@0:   beta = zeros(ss, ss);
Daniel@0:   P1sum = zeros(ss, ss);
Daniel@0:   x1sum = zeros(ss, 1);
Daniel@0:   loglik = 0;
Daniel@0:   
Daniel@0:   for ex = 1:N
Daniel@0:     y = data{ex};
Daniel@0:     T = length(y);
Daniel@0:     [beta_t, gamma_t, delta_t, gamma1_t, gamma2_t, x1, V1, loglik_t] = ...
Daniel@0: 	Estep(y, A, C, Q, R, initx, initV, ARmode);
Daniel@0:     beta = beta + beta_t;
Daniel@0:     gamma = gamma + gamma_t;
Daniel@0:     delta = delta + delta_t;
Daniel@0:     gamma1 = gamma1 + gamma1_t;
Daniel@0:     gamma2 = gamma2 + gamma2_t;
Daniel@0:     P1sum = P1sum + V1 + x1*x1';
Daniel@0:     x1sum = x1sum + x1;
Daniel@0:     %fprintf(1, 'example %d, ll/T %5.3f\n', ex, loglik_t/T);
Daniel@0:     loglik = loglik + loglik_t;
Daniel@0:   end
Daniel@0:   LL = [LL loglik];
Daniel@0:   if verbose, fprintf(1, 'iteration %d, loglik = %f\n', num_iter, loglik); end
Daniel@0:   %fprintf(1, 'iteration %d, loglik/NT = %f\n', num_iter, loglik/Tsum);
Daniel@0:   num_iter =  num_iter + 1;
Daniel@0:   
Daniel@0:   %%% M step
Daniel@0:   
Daniel@0:   % Tsum =  N*T
Daniel@0:   % Tsum1 = N*(T-1);
Daniel@0:   Tsum1 = Tsum - N;
Daniel@0:   A = beta * inv(gamma1);
Daniel@0:   %A = (gamma1' \ beta')';
Daniel@0:   Q = (gamma2 - A*beta') / Tsum1;
Daniel@0:   if diagQ
Daniel@0:     Q = diag(diag(Q));
Daniel@0:   end
Daniel@0:   if ~ARmode
Daniel@0:     C = delta * inv(gamma);
Daniel@0:     %C = (gamma' \ delta')';
Daniel@0:     R = (alpha - C*delta') / Tsum;
Daniel@0:     if diagR
Daniel@0:       R = diag(diag(R));
Daniel@0:     end
Daniel@0:   end
Daniel@0:   initx = x1sum / N;
Daniel@0:   initV = P1sum/N - initx*initx';
Daniel@0: 
Daniel@0:   if ~isempty(constr_fun)
Daniel@0:     [A,C,Q,R,initx,initV] = feval(constr_fun, A, C, Q, R, initx, initV, varargin{:});
Daniel@0:   end
Daniel@0:   
Daniel@0:   converged = em_converged(loglik, previous_loglik, thresh);
Daniel@0:   previous_loglik = loglik;
Daniel@0: end
Daniel@0: 
Daniel@0: 
Daniel@0: 
Daniel@0: %%%%%%%%%
Daniel@0: 
Daniel@0: function [beta, gamma, delta, gamma1, gamma2, x1, V1, loglik] = ...
Daniel@0:     Estep(y, A, C, Q, R, initx, initV, ARmode)
Daniel@0: %
Daniel@0: % Compute the (expected) sufficient statistics for a single Kalman filter sequence.
Daniel@0: %
Daniel@0: 
Daniel@0: [os T] = size(y);
Daniel@0: ss = length(A);
Daniel@0: 
Daniel@0: if ARmode
Daniel@0:   xsmooth = y;
Daniel@0:   Vsmooth = zeros(ss, ss, T); % no uncertainty about the hidden states
Daniel@0:   VVsmooth = zeros(ss, ss, T);
Daniel@0:   loglik = 0;
Daniel@0: else
Daniel@0:   [xsmooth, Vsmooth, VVsmooth, loglik] = kalman_smoother(y, A, C, Q, R, initx, initV);
Daniel@0: end
Daniel@0: 
Daniel@0: delta = zeros(os, ss);
Daniel@0: gamma = zeros(ss, ss);
Daniel@0: beta = zeros(ss, ss);
Daniel@0: for t=1:T
Daniel@0:   delta = delta + y(:,t)*xsmooth(:,t)';
Daniel@0:   gamma = gamma + xsmooth(:,t)*xsmooth(:,t)' + Vsmooth(:,:,t);
Daniel@0:   if t>1 beta = beta + xsmooth(:,t)*xsmooth(:,t-1)' + VVsmooth(:,:,t); end
Daniel@0: end
Daniel@0: gamma1 = gamma - xsmooth(:,T)*xsmooth(:,T)' - Vsmooth(:,:,T);
Daniel@0: gamma2 = gamma - xsmooth(:,1)*xsmooth(:,1)' - Vsmooth(:,:,1);
Daniel@0: 
Daniel@0: x1 = xsmooth(:,1);
Daniel@0: V1 = Vsmooth(:,:,1);
Daniel@0: 
Daniel@0: 
Daniel@0: