|
wolffd@0
|
1 function CPD = hhmmF_CPD(bnet, self, Qself, Fbelow, varargin)
|
|
wolffd@0
|
2 % HHMMF_CPD Make the CPD for an F node in a hierarchical HMM
|
|
wolffd@0
|
3 % CPD = hhmmF_CPD(bnet, self, Qself, Fbelow, ...)
|
|
wolffd@0
|
4 %
|
|
wolffd@0
|
5 % Qps
|
|
wolffd@0
|
6 % \
|
|
wolffd@0
|
7 % \
|
|
wolffd@0
|
8 % Fself
|
|
wolffd@0
|
9 % / |
|
|
wolffd@0
|
10 % / |
|
|
wolffd@0
|
11 % Qself Fbelow
|
|
wolffd@0
|
12 %
|
|
wolffd@0
|
13 % We assume nodes are ordered (numbered) as follows: Qps, Q, Fbelow, F
|
|
wolffd@0
|
14 % All nodes numbers should be from slice 1.
|
|
wolffd@0
|
15 %
|
|
wolffd@0
|
16 % If Fbelow if missing, this becomes a regular tabular_CPD.
|
|
wolffd@0
|
17 % Qps may be omitted.
|
|
wolffd@0
|
18 %
|
|
wolffd@0
|
19 % optional args [defaults]
|
|
wolffd@0
|
20 %
|
|
wolffd@0
|
21 % Qps - node numbers.
|
|
wolffd@0
|
22 % termprob - termprob(k,i,2) = prob finishing given Q(d)=i and Q(1:d-1)=k [ finish in last state wp 0.9]
|
|
wolffd@0
|
23 %
|
|
wolffd@0
|
24 % hhmmF_CPD is a subclass of tabular_CPD so we inherit inference methods like CPD_to_pot, etc.
|
|
wolffd@0
|
25 %
|
|
wolffd@0
|
26 % We create an isolated tabular_CPD with no F parent to learn termprob
|
|
wolffd@0
|
27 % so we can avail of e.g., entropic or Dirichlet priors.
|
|
wolffd@0
|
28 %
|
|
wolffd@0
|
29 % For details, see "Linear-time inference in hierarchical HMMs", Murphy and Paskin, NIPS'01.
|
|
wolffd@0
|
30
|
|
wolffd@0
|
31
|
|
wolffd@0
|
32
|
|
wolffd@0
|
33 Qps = [];
|
|
wolffd@0
|
34 % get parents
|
|
wolffd@0
|
35 for i=1:2:length(varargin)
|
|
wolffd@0
|
36 switch varargin{i},
|
|
wolffd@0
|
37 case 'Qps', Qps = varargin{i+1};
|
|
wolffd@0
|
38 end
|
|
wolffd@0
|
39 end
|
|
wolffd@0
|
40
|
|
wolffd@0
|
41 ns = bnet.node_sizes(:);
|
|
wolffd@0
|
42 Qsz = ns(Qself);
|
|
wolffd@0
|
43 Qpsz = prod(ns(Qps));
|
|
wolffd@0
|
44 CPD.Qsz = Qsz;
|
|
wolffd@0
|
45 CPD.Qpsz = Qpsz;
|
|
wolffd@0
|
46
|
|
wolffd@0
|
47 ps = parents(bnet.dag, self);
|
|
wolffd@0
|
48 CPD.Fbelow_ndx = find_equiv_posns(Fbelow, ps);
|
|
wolffd@0
|
49 CPD.Qps_ndx = find_equiv_posns(Qps, ps);
|
|
wolffd@0
|
50 CPD.Qself_ndx = find_equiv_posns(Qself, ps);
|
|
wolffd@0
|
51
|
|
wolffd@0
|
52 % set default arguments
|
|
wolffd@0
|
53 p = 0.9;
|
|
wolffd@0
|
54 %termprob(k,i,t) Might terminate if i=Qsz; will not terminate if i<Qsz
|
|
wolffd@0
|
55 termprob = zeros(Qpsz, Qsz, 2);
|
|
wolffd@0
|
56 termprob(:, Qsz, 2) = p;
|
|
wolffd@0
|
57 termprob(:, Qsz, 1) = 1-p;
|
|
wolffd@0
|
58 termprob(:, 1:(Qsz-1), 1) = 1;
|
|
wolffd@0
|
59
|
|
wolffd@0
|
60 for i=1:2:length(varargin)
|
|
wolffd@0
|
61 switch varargin{i},
|
|
wolffd@0
|
62 case 'termprob', termprob = varargin{i+1};
|
|
wolffd@0
|
63 end
|
|
wolffd@0
|
64 end
|
|
wolffd@0
|
65
|
|
wolffd@0
|
66 CPD.sub_CPD_term = mk_isolated_tabular_CPD([Qpsz Qsz 2], {'CPT', termprob});
|
|
wolffd@0
|
67 S = struct(CPD.sub_CPD_term);
|
|
wolffd@0
|
68 CPD.termprob = S.CPT;
|
|
wolffd@0
|
69
|
|
wolffd@0
|
70 CPD = class(CPD, 'hhmmF_CPD', tabular_CPD(bnet, self));
|
|
wolffd@0
|
71
|
|
wolffd@0
|
72 CPD = update_CPT(CPD);
|
|
wolffd@0
|
73
|
