wolffd@0: % Computing most probable explanation.
wolffd@0: 
wolffd@0: % If you don't break ties consistently, loopy can give wrong mpe
wolffd@0: % even though the graph has no cycles, and even though the max-marginals are the same.
wolffd@0: % This example was contributed by Wentau Yih <wtyih@yahoo.com> 29 Jan 02.
wolffd@0: 
wolffd@0: % define loop-free graph structure (all edges point down)
wolffd@0: %
wolffd@0: % Xe1   Xe2
wolffd@0: %  |    |
wolffd@0: %  E1   E2
wolffd@0: %    \ /
wolffd@0: %     R1
wolffd@0: %     |
wolffd@0: %    Xr1
wolffd@0: 
wolffd@0: N = 6;
wolffd@0: dag = zeros(N,N);
wolffd@0: Xe1 = 1; Xe2 = 2; E1 = 3; E2 = 4; R1 = 5; Xr1 = 6;
wolffd@0: dag(Xe1, E1) = 1;
wolffd@0: dag(Xe2, E2) = 1;
wolffd@0: dag([E1 E2], R1) = 1;
wolffd@0: dag(R1, Xr1) = 1;
wolffd@0: 
wolffd@0: node_sizes = [ 1 1 2 2 2 1 ];
wolffd@0: 
wolffd@0: % create BN
wolffd@0: 
wolffd@0: bnet = mk_bnet(dag, node_sizes, 'observed', [Xe1 Xe2 Xr1]);
wolffd@0: 
wolffd@0: % fill in CPT
wolffd@0: 
wolffd@0: bnet.CPD{Xe1} = tabular_CPD(bnet, Xe1, [1]);
wolffd@0: bnet.CPD{Xe2} = tabular_CPD(bnet, Xe2, [1]);
wolffd@0: bnet.CPD{E1} = tabular_CPD(bnet, E1, [0.2 0.8]);
wolffd@0: bnet.CPD{E2} = tabular_CPD(bnet, E2, [0.3 0.7]);
wolffd@0: bnet.CPD{R1} = tabular_CPD(bnet, R1, [1 1 1 0.8 0 0 0 0.2]);
wolffd@0: bnet.CPD{Xr1} = tabular_CPD(bnet, Xr1, [0.15 0.85]);
wolffd@0: 
wolffd@0: clear engine;
wolffd@0: engine{1} = belprop_inf_engine(bnet);
wolffd@0: engine{2} = jtree_inf_engine(bnet);
wolffd@0: engine{3} = global_joint_inf_engine(bnet);
wolffd@0: engine{4} = var_elim_inf_engine(bnet);
wolffd@0: 
wolffd@0: evidence = cell(1,N);
wolffd@0: evidence{Xe1} = 1;  evidence{Xe2} = 1;  evidence{Xr1} = 1;
wolffd@0: 
wolffd@0: mpe = find_mpe(engine{1}, evidence, 'break_ties', 0) % gives wrong results
wolffd@0: mpe = find_mpe(engine{1}, evidence)
wolffd@0: for i=2:4
wolffd@0:   mpe = find_mpe(engine{i}, evidence)
wolffd@0: end