dbn_to_hmm.m

Bayesian网络工具箱.

function [startprob, transprob, obsprob] = dbn_to_hmm(bnet)
% DBN_TO_HMM % Convert DBN params to HMM params
% [startprob, transprob, obsprob] = dbn_to_hmm(bnet, onodes)
% startprob(i)
% transprob(i,j)
% obsprob{k}(i,o) if k'th observed node is discrete
% obsprob{k}.mu(:,i), .Sigma(:,:,i) if k'th observed node is Gaussian
%

% Called by constructor and by update_engine

ss = length(bnet.intra);
onodes = bnet.observed;
hnodes = mysetdiff(1:ss, onodes);
evidence = cell(ss, 2);
ns = bnet.node_sizes;
Q = prod(ns(hnodes));
tmp = dpot_to_table(compute_joint_pot(bnet, hnodes, evidence));
startprob = reshape(tmp, Q, 1);

tmp = dpot_to_table(compute_joint_pot(bnet, hnodes+ss, evidence, [hnodes hnodes+ss]));
transprob = mk_stochastic(reshape(tmp, Q, Q));

obsprob = cell(1, length(onodes));
for i=1:length(onodes)
  if bnet.ar_hmm
    CPD = struct(bnet.CPD{bnet.equiv_class(onodes(i))});
    obsprob{i}.mu0 = CPD.mean;
    obsprob{i}.Sigma0 = CPD.cov;

    CPD = struct(bnet.CPD{bnet.equiv_class(onodes(i)+ss)});
    obsprob{i}.mu = CPD.mean;
    obsprob{i}.Sigma = CPD.cov;
    obsprob{i}.W = CPD.weights;
  else
    jpot = pot_to_marginal(compute_joint_pot(bnet, onodes(i), evidence, [hnodes onodes(i)]));
    if myismember(onodes(i), bnet.dnodes)
      O = ns(onodes(i));
      tmp = reshape(jpot.T, Q, O);
      assert(is_stochastic(tmp));
      obsprob{i} = tmp;
      %obsprob{i} = mk_stochastic(reshape(jpot.T, Q, O));
    else
      %obsprob{i} = jpot;
      obsprob{i}.mu = jpot.mu;
      C = jpot.Sigma;
      obsprob{i}.Sigma = C;
      obsprob{i}.inv_Sigma = inv(C);
      d = length(jpot.mu);
      obsprob{i}.denom = (2*pi)^(d/2)*sqrt(abs(det(C)));
    end
  end
end