wolffd@0: function [startprob, transprob, obsprob] = dbn_to_hmm(bnet)
wolffd@0: % DBN_TO_HMM % Convert DBN params to HMM params
wolffd@0: % [startprob, transprob, obsprob] = dbn_to_hmm(bnet, onodes)
wolffd@0: % startprob(i)
wolffd@0: % transprob(i,j)
wolffd@0: % obsprob{k}.big_CPT(i,o) if k'th observed node is discrete
wolffd@0: % obsprob{k}.big_mu(:,i), .big_Sigma(:,:,i) if k'th observed node is Gaussian
wolffd@0: % Big means the domain contains all the hidden discrete nodes, not just the parents.
wolffd@0: 
wolffd@0: % Called by constructor and by update_engine
wolffd@0: 
wolffd@0: ss = length(bnet.intra);
wolffd@0: onodes = bnet.observed;
wolffd@0: hnodes = mysetdiff(1:ss, onodes);
wolffd@0: evidence = cell(ss, 2);
wolffd@0: ns = bnet.node_sizes(:);
wolffd@0: Qh = prod(ns(hnodes));
wolffd@0: tmp = dpot_to_table(compute_joint_pot(bnet, hnodes, evidence));
wolffd@0: startprob = reshape(tmp, Qh, 1);
wolffd@0: 
wolffd@0: tmp = dpot_to_table(compute_joint_pot(bnet, hnodes+ss, evidence, [hnodes hnodes+ss]));
wolffd@0: transprob = mk_stochastic(reshape(tmp, Qh, Qh));
wolffd@0: 
wolffd@0: % P(o|ps) is used by mk_hmm_obs_lik_vec for a single time slice
wolffd@0: % P(o|h) (the big version), where h = all hidden nodes, is used by enter_evidence
wolffd@0: 
wolffd@0: obsprob = cell(1, length(onodes));
wolffd@0: for i=1:length(onodes)
wolffd@0:   o = onodes(i);
wolffd@0:   if bnet.auto_regressive(o)
wolffd@0:     % We assume the parents of this node are all the hidden nodes in the slice,
wolffd@0:     % so the params already are "big". Also, we assume we regress only on our old selves.
wolffd@0:     % slice 1
wolffd@0:     e = bnet.equiv_class(o);
wolffd@0:     CPD = struct(bnet.CPD{e});
wolffd@0:     O = ns(o);
wolffd@0:     ps = bnet.parents{o};
wolffd@0:     Qps = prod(ns(ps));
wolffd@0:     obsprob{i}.big_mu0 = reshape(CPD.mean, [O Qps]);
wolffd@0:     obsprob{i}.big_Sigma0 = reshape(CPD.cov, [O O Qps]);
wolffd@0: 
wolffd@0:     % slice t>1
wolffd@0:     e = bnet.equiv_class(o+ss);
wolffd@0:     CPD = struct(bnet.CPD{e});
wolffd@0:     O = ns(o);
wolffd@0:     dps = mysetdiff(bnet.parents{o+ss}, o);
wolffd@0:     Qdps = prod(ns(dps));
wolffd@0:     obsprob{i}.big_mu = reshape(CPD.mean, [O Qdps]);
wolffd@0:     obsprob{i}.big_Sigma = reshape(CPD.cov, [O O Qdps]);
wolffd@0:     obsprob{i}.big_W = reshape(CPD.weights, [O O Qdps]);
wolffd@0:   else
wolffd@0:     e = bnet.equiv_class(o+ss);
wolffd@0:     CPD = struct(bnet.CPD{e});
wolffd@0:     O = ns(o);
wolffd@0:     ps = bnet.parents{o};
wolffd@0:     Qps = prod(ns(ps));
wolffd@0:     % We make a big potential, replicating the params if necessary
wolffd@0:     % e.g., for a 2 chain coupled HMM, mu(:,Q1) becomes mu(:,Q1,Q2)
wolffd@0:     bigpot = pot_to_marginal(compute_joint_pot(bnet, onodes(i), evidence, [hnodes onodes(i)]));
wolffd@0: 
wolffd@0:     if myismember(o, bnet.dnodes)
wolffd@0:       obsprob{i}.CPT = reshape(CPD.CPT, [Qps O]);
wolffd@0:       obsprob{i}.big_CPT = reshape(bigpot.T, Qh, O); 
wolffd@0:     else
wolffd@0:       obsprob{i}.big_mu = bigpot.mu;
wolffd@0:       obsprob{i}.big_Sigma = bigpot.Sigma;
wolffd@0:       
wolffd@0:       if 1
wolffd@0:       obsprob{i}.mu = reshape(CPD.mean, [O Qps]);
wolffd@0:       C = reshape(CPD.cov, [O O Qps]);
wolffd@0:       obsprob{i}.Sigma = C;
wolffd@0:       d = size(obsprob{i}.mu, 1);
wolffd@0:       for j=1:Qps
wolffd@0: 	obsprob{i}.inv_Sigma(:,:,j) = inv(C(:,:,j));
wolffd@0: 	obsprob{i}.denom(j) = (2*pi)^(d/2)*sqrt(abs(det(C(:,:,j))));
wolffd@0:       end
wolffd@0:       end
wolffd@0:       
wolffd@0:     end % if discrete
wolffd@0:   end % if ar
wolffd@0: end % for