wolffd@0
|
1 % Compare various inference engines on the following network (from Jensen (1996) p84 fig 4.17)
|
wolffd@0
|
2 % 1
|
wolffd@0
|
3 % / | \
|
wolffd@0
|
4 % 2 3 4
|
wolffd@0
|
5 % | | |
|
wolffd@0
|
6 % 5 6 7
|
wolffd@0
|
7 % \/ \/
|
wolffd@0
|
8 % 8 9
|
wolffd@0
|
9 % where all arcs point downwards
|
wolffd@0
|
10 seed = 0;
|
wolffd@0
|
11 rand('state', seed);
|
wolffd@0
|
12 randn('state', seed);
|
wolffd@0
|
13
|
wolffd@0
|
14 N = 9;
|
wolffd@0
|
15 dag = zeros(N,N);
|
wolffd@0
|
16 dag(1,2)=1; dag(1,3)=1; dag(1,4)=1;
|
wolffd@0
|
17 dag(2,5)=1; dag(3,6)=1; dag(4,7)=1;
|
wolffd@0
|
18 dag(5,8)=1; dag(6,8)=1; dag(6,9)=1; dag(7,9) = 1;
|
wolffd@0
|
19
|
wolffd@0
|
20 dnodes = 1:N;
|
wolffd@0
|
21 false = 1; true = 2;
|
wolffd@0
|
22 ns = 2*ones(1,N); % binary nodes
|
wolffd@0
|
23
|
wolffd@0
|
24 onodes = [2 4];
|
wolffd@0
|
25 bnet = mk_bnet(dag, ns, 'observed', onodes);
|
wolffd@0
|
26 % use random params
|
wolffd@0
|
27 for i=1:N
|
wolffd@0
|
28 bnet.CPD{i} = tabular_CPD(bnet, i);
|
wolffd@0
|
29 end
|
wolffd@0
|
30
|
wolffd@0
|
31 %USEC = exist('@jtree_C_inf_engine/collect_evidence','file');
|
wolffd@0
|
32 query = [3];
|
wolffd@0
|
33 engine = {};
|
wolffd@0
|
34 engine{end+1} = jtree_inf_engine(bnet);
|
wolffd@0
|
35 engine{end+1} = jtree_sparse_inf_engine(bnet);
|
wolffd@0
|
36 %engine{end+1} = jtree_ndx_inf_engine(bnet, 'ndx_type', 'SD');
|
wolffd@0
|
37 %engine{end+1} = jtree_ndx_inf_engine(bnet, 'ndx_type', 'B');
|
wolffd@0
|
38 %engine{end+1} = jtree_ndx_inf_engine(bnet, 'ndx_type', 'D');
|
wolffd@0
|
39 %if USEC, engine{end+1} = jtree_C_inf_engine(bnet); end
|
wolffd@0
|
40 %engine{end+1} = var_elim_inf_engine(bnet);
|
wolffd@0
|
41 %engine{end+1} = enumerative_inf_engine(bnet);
|
wolffd@0
|
42 %engine{end+1} = jtree_onepass_inf_engine(bnet, query, onodes);
|
wolffd@0
|
43
|
wolffd@0
|
44 maximize = 0; % jtree_ndx crashes on max-prop
|
wolffd@0
|
45 [err, time] = cmp_inference_static(bnet, engine, 'maximize', maximize);
|
wolffd@0
|
46
|