diff toolboxes/FullBNT-1.0.7/bnt/examples/dynamic/mk_chmm.m @ 0:e9a9cd732c1e tip

first hg version after svn
author wolffd
date Tue, 10 Feb 2015 15:05:51 +0000
parents
children
line wrap: on
line diff
--- /dev/null	Thu Jan 01 00:00:00 1970 +0000
+++ b/toolboxes/FullBNT-1.0.7/bnt/examples/dynamic/mk_chmm.m	Tue Feb 10 15:05:51 2015 +0000
@@ -0,0 +1,103 @@
+function bnet = mk_chmm(N, Q, Y, discrete_obs, coupled, CPD)
+% MK_CHMM Make a coupled Hidden Markov Model
+%
+% There are N hidden nodes, each connected to itself and its two nearest neighbors in the next
+% slice (apart from the edges, where there is 1 nearest neighbor).
+%
+% Example: If N = 3, the hidden backbone is as follows, where all arrows point to the righ+t
+%
+% X1--X2
+%   \/ 
+%   /\
+% X2--X2
+%   \/ 
+%   /\
+% X3--X3
+%
+% Each hidden node has a "private" observed child (not shown).
+%
+% BNET = MK_CHMM(N, Q, Y)
+% Each hidden node is discrete and has Q values.
+% Each observed node is a Gaussian vector of length Y.
+%
+% BNET = MK_CHMM(N, Q, Y, DISCRETE_OBS)
+% If discrete_obs = 1, the observations are discrete (values in {1, .., Y}).
+%
+% BNET = MK_CHMM(N, Q, Y, DISCRETE_OBS, COUPLED)
+% If coupled = 0, the chains are not coupled, i.e., we make N parallel HMMs.
+%
+% BNET = MK_CHMM(N, Q, Y, DISCRETE_OBS, COUPLED, CPDs)
+% means use the specified CPD structures instead of creating random params.
+%  CPD{i}.CPT, i=1:N specifies the prior
+%  CPD{i}.CPT, i=2N+1:3N specifies the transition model
+%  CPD{i}.mean, CPD{i}.cov, i=N+1:2N specifies the observation model if Gaussian
+%  CPD{i}.CPT, i=N+1:2N if discrete
+
+
+if nargin < 2, Q = 2; end
+if nargin < 3, Y = 1; end
+if nargin < 4, discrete_obs = 0; end
+if nargin < 5, coupled = 1; end
+if nargin < 6, rnd = 1; else rnd = 0; end
+  
+ss = N*2;
+hnodes = 1:N;
+onodes = (1:N)+N;
+
+intra = zeros(ss);
+for i=1:N
+  intra(hnodes(i), onodes(i))=1;
+end
+
+inter = zeros(ss);
+if coupled
+  for i=1:N
+    inter(i, max(i-1,1):min(i+1,N))=1;
+  end
+else
+  inter(1:N, 1:N) = eye(N);
+end  
+
+ns = [Q*ones(1,N) Y*ones(1,N)]; 
+
+eclass1 = [hnodes onodes];
+eclass2 = [hnodes+ss onodes];
+if discrete_obs
+  dnodes = 1:ss;
+else
+  dnodes = hnodes;
+end
+bnet = mk_dbn(intra, inter, ns, 'discrete', dnodes, 'eclass1', eclass1, 'eclass2', eclass2, ...
+	      'observed', onodes);
+
+if rnd
+  for i=hnodes(:)'
+    bnet.CPD{i} = tabular_CPD(bnet, i);
+  end
+  for i=onodes(:)'
+    if discrete_obs
+      bnet.CPD{i} = tabular_CPD(bnet, i);
+    else
+      bnet.CPD{i} = gaussian_CPD(bnet, i);
+    end
+  end
+  for i=hnodes(:)'+ss
+    bnet.CPD{i} = tabular_CPD(bnet, i);
+  end
+else
+  for i=hnodes(:)'
+    bnet.CPD{i} = tabular_CPD(bnet, i, CPD{i}.CPT);
+  end
+  for i=onodes(:)'
+    if discrete_obs
+      bnet.CPD{i} = tabular_CPD(bnet, i, CPD{i}.CPT);
+    else
+      bnet.CPD{i} = gaussian_CPD(bnet, i, CPD{i}.mean, CPD{i}.cov);
+    end
+  end
+  for i=hnodes(:)'+ss
+    bnet.CPD{i} = tabular_CPD(bnet, i, CPD{i}.CPT);
+  end
+end
+
+