|
wolffd@0
|
1 function [engine, loglik] = enter_evidence(engine, evidence, varargin)
|
|
wolffd@0
|
2 % ENTER_EVIDENCE Add the specified evidence to the network (kalman)
|
|
wolffd@0
|
3 % [engine, loglik] = enter_evidence(engine, evidence, ...)
|
|
wolffd@0
|
4 %
|
|
wolffd@0
|
5 % evidence{i,t} = [] if if X(i,t) is hidden, and otherwise contains its observed value (scalar or column vector)
|
|
wolffd@0
|
6 %
|
|
wolffd@0
|
7 % The following optional arguments can be specified in the form of name/value pairs:
|
|
wolffd@0
|
8 % [default value in brackets]
|
|
wolffd@0
|
9 %
|
|
wolffd@0
|
10 % maximize - if 1, does max-product (same as sum-product for Gaussians!), else sum-product [0]
|
|
wolffd@0
|
11 % filter - if 1, do filtering, else smoothing [0]
|
|
wolffd@0
|
12 %
|
|
wolffd@0
|
13 % e.g., engine = enter_evidence(engine, ev, 'maximize', 1)
|
|
wolffd@0
|
14
|
|
wolffd@0
|
15 maximize = 0;
|
|
wolffd@0
|
16 filter = 0;
|
|
wolffd@0
|
17
|
|
wolffd@0
|
18 % parse optional params
|
|
wolffd@0
|
19 args = varargin;
|
|
wolffd@0
|
20 nargs = length(args);
|
|
wolffd@0
|
21 if nargs > 0
|
|
wolffd@0
|
22 for i=1:2:nargs
|
|
wolffd@0
|
23 switch args{i},
|
|
wolffd@0
|
24 case 'maximize', maximize = args{i+1};
|
|
wolffd@0
|
25 case 'filter', filter = args{i+1};
|
|
wolffd@0
|
26 otherwise,
|
|
wolffd@0
|
27 error(['invalid argument name ' args{i}]);
|
|
wolffd@0
|
28 end
|
|
wolffd@0
|
29 end
|
|
wolffd@0
|
30 end
|
|
wolffd@0
|
31
|
|
wolffd@0
|
32 assert(~maximize);
|
|
wolffd@0
|
33
|
|
wolffd@0
|
34 bnet = bnet_from_engine(engine);
|
|
wolffd@0
|
35 n = length(bnet.intra);
|
|
wolffd@0
|
36 onodes = bnet.observed;
|
|
wolffd@0
|
37 hnodes = mysetdiff(1:n, onodes);
|
|
wolffd@0
|
38 T = size(evidence, 2);
|
|
wolffd@0
|
39 ns = bnet.node_sizes;
|
|
wolffd@0
|
40 O = sum(ns(onodes));
|
|
wolffd@0
|
41 data = reshape(cat(1, evidence{onodes,:}), [O T]);
|
|
wolffd@0
|
42
|
|
wolffd@0
|
43 A = engine.trans_mat;
|
|
wolffd@0
|
44 C = engine.obs_mat;
|
|
wolffd@0
|
45 Q = engine.trans_cov;
|
|
wolffd@0
|
46 R = engine.obs_cov;
|
|
wolffd@0
|
47 init_x = engine.init_state;
|
|
wolffd@0
|
48 init_V = engine.init_cov;
|
|
wolffd@0
|
49
|
|
wolffd@0
|
50 if filter
|
|
wolffd@0
|
51 [x, V, VV, loglik] = kalman_filter(data, A, C, Q, R, init_x, init_V);
|
|
wolffd@0
|
52 else
|
|
wolffd@0
|
53 [x, V, VV, loglik] = kalman_smoother(data, A, C, Q, R, init_x, init_V);
|
|
wolffd@0
|
54 end
|
|
wolffd@0
|
55
|
|
wolffd@0
|
56
|
|
wolffd@0
|
57 % Wrap the posterior inside a potential, so it can be marginalized easily
|
|
wolffd@0
|
58 engine.one_slice_marginal = cell(1,T);
|
|
wolffd@0
|
59 engine.two_slice_marginal = cell(1,T);
|
|
wolffd@0
|
60 ns(onodes) = 0;
|
|
wolffd@0
|
61 ns(onodes+n) = 0;
|
|
wolffd@0
|
62 ss = length(bnet.intra);
|
|
wolffd@0
|
63 for t=1:T
|
|
wolffd@0
|
64 dom = (1:n);
|
|
wolffd@0
|
65 engine.one_slice_marginal{t} = mpot(dom+(t-1)*ss, ns(dom), 1, x(:,t), V(:,:,t));
|
|
wolffd@0
|
66 end
|
|
wolffd@0
|
67 % for t=1:T-1
|
|
wolffd@0
|
68 % dom = (1:(2*n));
|
|
wolffd@0
|
69 % mu = [x(:,t); x(:,t)];
|
|
wolffd@0
|
70 % Sigma = [V(:,:,t) VV(:,:,t+1)';
|
|
wolffd@0
|
71 % VV(:,:,t+1) V(:,:,t+1)];
|
|
wolffd@0
|
72 % engine.two_slice_marginal{t} = mpot(dom+(t-1)*ss, ns(dom), 1, mu, Sigma);
|
|
wolffd@0
|
73 % end
|
|
wolffd@0
|
74 for t=2:T
|
|
wolffd@0
|
75 %dom = (1:(2*n));
|
|
wolffd@0
|
76 current_slice = hnodes;
|
|
wolffd@0
|
77 next_slice = hnodes + ss;
|
|
wolffd@0
|
78 dom = [current_slice next_slice];
|
|
wolffd@0
|
79 mu = [x(:,t-1); x(:,t)];
|
|
wolffd@0
|
80 Sigma = [V(:,:,t-1) VV(:,:,t)';
|
|
wolffd@0
|
81 VV(:,:,t) V(:,:,t)];
|
|
wolffd@0
|
82 engine.two_slice_marginal{t-1} = mpot(dom+(t-2)*ss, ns(dom), 1, mu, Sigma);
|
|
wolffd@0
|
83 end
|
