Mercurial > hg > camir-aes2014
comparison toolboxes/FullBNT-1.0.7/bnt/CPDs/@gaussian_CPD/CPD_to_lambda_msg.m @ 0:e9a9cd732c1e tip
first hg version after svn
author | wolffd |
---|---|
date | Tue, 10 Feb 2015 15:05:51 +0000 |
parents | |
children |
comparison
equal
deleted
inserted
replaced
-1:000000000000 | 0:e9a9cd732c1e |
---|---|
1 function lam_msg = CPD_to_lambda_msg(CPD, msg_type, n, ps, msg, p, evidence) | |
2 % CPD_TO_LAMBDA_MSG Compute lambda message (gaussian) | |
3 % lam_msg = compute_lambda_msg(CPD, msg_type, n, ps, msg, p, evidence) | |
4 % Pearl p183 eq 4.52 | |
5 | |
6 switch msg_type | |
7 case 'd', | |
8 error('gaussian_CPD can''t create discrete msgs') | |
9 case 'g', | |
10 cps = ps(CPD.cps); | |
11 cpsizes = CPD.sizes(CPD.cps); | |
12 self_size = CPD.sizes(end); | |
13 i = find_equiv_posns(p, cps); % p is n's i'th cts parent | |
14 psz = cpsizes(i); | |
15 if all(msg{n}.lambda.precision == 0) % no info to send on | |
16 lam_msg.precision = zeros(psz, psz); | |
17 lam_msg.info_state = zeros(psz, 1); | |
18 return; | |
19 end | |
20 [m, Q, W] = gaussian_CPD_params_given_dps(CPD, [ps n], evidence); | |
21 Bmu = m; | |
22 BSigma = Q; | |
23 for k=1:length(cps) % only get pi msgs from cts parents | |
24 pk = cps(k); | |
25 if pk ~= p | |
26 %bk = block(k, cpsizes); | |
27 bk = CPD.cps_block_ndx{k}; | |
28 Bk = W(:, bk); | |
29 m = msg{n}.pi_from_parent{k}; | |
30 BSigma = BSigma + Bk * m.Sigma * Bk'; | |
31 Bmu = Bmu + Bk * m.mu; | |
32 end | |
33 end | |
34 % BSigma = Q + sum_{k \neq i} B_k Sigma_k B_k' | |
35 %bi = block(i, cpsizes); | |
36 bi = CPD.cps_block_ndx{i}; | |
37 Bi = W(:,bi); | |
38 P = msg{n}.lambda.precision; | |
39 if (rcond(P) > 1e-3) | isinf(P) | |
40 if isinf(P) % Y is observed | |
41 Sigma_lambda = zeros(self_size, self_size); % infinite precision => 0 variance | |
42 mu_lambda = msg{n}.lambda.mu; % observed_value; | |
43 else | |
44 Sigma_lambda = inv(P); | |
45 mu_lambda = Sigma_lambda * msg{n}.lambda.info_state; | |
46 end | |
47 C = inv(Sigma_lambda + BSigma); | |
48 lam_msg.precision = Bi' * C * Bi; | |
49 lam_msg.info_state = Bi' * C * (mu_lambda - Bmu); | |
50 else | |
51 % method that uses matrix inversion lemma to avoid inverting P | |
52 A = inv(P + inv(BSigma)); | |
53 C = P - P*A*P; | |
54 lam_msg.precision = Bi' * C * Bi; | |
55 D = eye(self_size) - P*A; | |
56 z = msg{n}.lambda.info_state; | |
57 lam_msg.info_state = Bi' * (D*z - D*P*Bmu); | |
58 end | |
59 end |