diff toolboxes/FullBNT-1.0.7/graph/cliques_to_strong_jtree.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/graph/cliques_to_strong_jtree.m	Tue Feb 10 15:05:51 2015 +0000
@@ -0,0 +1,114 @@
+function [jtree, root, cliques, B, w] = mk_strong_jtree(cliques, ns, elim_order, MTG)
+% MK_SRONG_JTREE Make a strong junction tree.
+% [jtree, root, cliques, B, w] = mk_strong_jtree(cliques, ns, elim_order, MTG)
+%
+% Here is a definition of a strong jtree from Jensen et al. 1994:
+% "A junction tree is said to be strong if it has at least one distinguished clique R,
+% called a strong root, s.t. for each pair (C1,C2) of adjacent cliques in the tree,
+% with C1 closer to R than C2, there exists and ordering of [the nodes below] C2
+% that respects [the partial order] and with the vertices of the separator C1 intersect C2
+% preceeding the vertices [below C2] of C2 \ C1."
+%          
+% For details, see
+% - Jensen, Jensen and Dittmer, "From influence diagrams to junction trees", UAI 94.
+%
+% MTG is the moralized, triangulated graph.
+% elim_order is the elimination ordering used to compute MTG.
+
+
+% Warning: this is a very naive implementation of the algorithm in Jensen et al.
+
+n = length(elim_order);
+alpha(elim_order) = n:-1:1;
+% alpha(u) = i if we eliminate u at step n-i+1
+% i.e., vertices with higher alpha numbers are eliminated before vertices with lower numbers.
+% e.g., from the Jensen et al paper
+% node a=1 eliminated at step 6, so alpha(a)=16-6+1=11.
+% alpha = [11 1 2 10 9 3 4 7 5 8 13 12 6 16 15 14]
+
+
+% We sort the cliques in order of increasing index. The index of a clique C is defined as follows.
+% Let lower = {u | alpha(u) < alpha(v)}, and
+% let v in C be the highest-numbered vertex s.t. the vertices in W = lower intersect C
+% have a common neighbor u in U, where U = lower \ C.
+% If such a v exists, define index(C) = alpha(v), otherwise, index(C) = 1.
+% Intuitively, index(C) is the step in the elimination process at which C disappears.
+
+num_cliques = length(cliques);
+index = zeros(1, num_cliques);
+for c = 1:num_cliques
+  C = cliques{c};
+  highest_num = -inf;
+  for vi = 1:length(C)
+    v = C(vi);
+    lower = find(alpha < alpha(v));
+    W = myintersect(lower, C);
+    U = mysetdiff(lower, C);
+    found = 0;
+    for ui=1:length(U)
+      u = U(ui);
+      if mysubset(W, neighbors(MTG, u))
+	found = 1;
+	break;
+      end
+    end
+    if found
+      if alpha(v) > highest_num
+	highest_num = alpha(v);
+      end
+    end
+  end
+  if highest_num == -inf
+    index(c) = 1;
+  else
+    index(c) = highest_num;
+  end
+end
+
+
+% Permute the cliques so that they are ordered according to index
+[dummy, clique_order] = sort(index);
+cliques = cliques(clique_order);
+
+w = zeros(num_cliques, 1); 
+B = sparse(num_cliques, 1);
+for i=1:num_cliques
+  B(i, cliques{i}) = 1;
+  w(i) = prod(ns(cliques{i}));
+end
+
+% Pearl p113 suggests ordering the cliques by rank of the highest vertex in each clique.
+% However, this will only work if we use maximum cardinality search.
+
+
+% Join up the cliques so that they satisfy the Running Intersection Property.
+% This states that, for all k > 1, S(k) subseteq C(j) for some j < k, where
+% S(k) = C(k) intersect (union_{i=1}^{k-1} C(i))
+jtree = sparse(num_cliques, num_cliques);
+for k=2:num_cliques
+  S = [];
+  for i=1:k-1
+    S = myunion(S, cliques{i});
+  end
+  S = myintersect(S, cliques{k});
+  found = 0;
+  for j=1:k-1
+    if mysubset(S, cliques{j})
+      found = 1;
+      break;
+    end
+  end
+  if ~found
+    disp(['RIP is violated for clique ' num2str(k)]);
+  end
+  jtree(k,j)=1;
+  jtree(j,k)=1;
+end
+
+% Pearl p113 suggests connecting Ci to a predecessor Cj (j < i) sharing
+% the highest number of vertices with Ci (i.e., the heaviest i-j edge
+% in the jgraph). However, this will only work if we use maximum cardinality search.
+
+root = 1;
+
+