|
wolffd@0
|
1 % Compute Viterbi path discrete HMM by different methods
|
|
wolffd@0
|
2
|
|
wolffd@0
|
3 intra = zeros(2);
|
|
wolffd@0
|
4 intra(1,2) = 1;
|
|
wolffd@0
|
5 inter = zeros(2);
|
|
wolffd@0
|
6 inter(1,1) = 1;
|
|
wolffd@0
|
7 n = 2;
|
|
wolffd@0
|
8
|
|
wolffd@0
|
9 Q = 2; % num hidden states
|
|
wolffd@0
|
10 O = 2; % num observable symbols
|
|
wolffd@0
|
11
|
|
wolffd@0
|
12 ns = [Q O];
|
|
wolffd@0
|
13 dnodes = 1:2;
|
|
wolffd@0
|
14 onodes = [2];
|
|
wolffd@0
|
15 eclass1 = [1 2];
|
|
wolffd@0
|
16 eclass2 = [3 2];
|
|
wolffd@0
|
17 bnet = mk_dbn(intra, inter, ns, 'discrete', dnodes, 'eclass1', eclass1, 'eclass2', eclass2, ...
|
|
wolffd@0
|
18 'observed', onodes);
|
|
wolffd@0
|
19
|
|
wolffd@0
|
20 for seed=1:10
|
|
wolffd@0
|
21 rand('state', seed);
|
|
wolffd@0
|
22 prior = normalise(rand(Q,1));
|
|
wolffd@0
|
23 transmat = mk_stochastic(rand(Q,Q));
|
|
wolffd@0
|
24 obsmat = mk_stochastic(rand(Q,O));
|
|
wolffd@0
|
25 bnet.CPD{1} = tabular_CPD(bnet, 1, prior);
|
|
wolffd@0
|
26 bnet.CPD{2} = tabular_CPD(bnet, 2, obsmat);
|
|
wolffd@0
|
27 bnet.CPD{3} = tabular_CPD(bnet, 3, transmat);
|
|
wolffd@0
|
28
|
|
wolffd@0
|
29
|
|
wolffd@0
|
30 % Create a sequence
|
|
wolffd@0
|
31 T = 5;
|
|
wolffd@0
|
32 ev = sample_dbn(bnet, T);
|
|
wolffd@0
|
33 evidence = cell(2,T);
|
|
wolffd@0
|
34 evidence(2,:) = ev(2,:); % extract observed component
|
|
wolffd@0
|
35 data = cell2num(ev(2,:));
|
|
wolffd@0
|
36
|
|
wolffd@0
|
37 %obslik = mk_dhmm_obs_lik(data, obsmat);
|
|
wolffd@0
|
38 obslik = multinomial_prob(data, obsmat);
|
|
wolffd@0
|
39 path = viterbi_path(prior, transmat, obslik);
|
|
wolffd@0
|
40
|
|
wolffd@0
|
41 engine = {};
|
|
wolffd@0
|
42 engine{end+1} = smoother_engine(jtree_2TBN_inf_engine(bnet));
|
|
wolffd@0
|
43
|
|
wolffd@0
|
44 mpe = find_mpe(engine{1}, evidence);
|
|
wolffd@0
|
45
|
|
wolffd@0
|
46 assert(isequal(cell2num(mpe(1,:)), path)) % extract values of hidden nodes
|
|
wolffd@0
|
47 end
|
