annotate toolboxes/FullBNT-1.0.7/bnt/examples/static/Belprop/gmux1.m @ 0:e9a9cd732c1e tip

first hg version after svn
author wolffd
date Tue, 10 Feb 2015 15:05:51 +0000
parents
children
rev   line source
wolffd@0 1 % Test gmux.
wolffd@0 2 % The following model, where Y is a gmux node,
wolffd@0 3 % and M is set to 1, should be equivalent to X1 -> Y
wolffd@0 4 %
wolffd@0 5 % X1 Xn M
wolffd@0 6 % \ | /
wolffd@0 7 % Y
wolffd@0 8
wolffd@0 9 n = 3;
wolffd@0 10 N = n+2;
wolffd@0 11 Xs = 1:n;
wolffd@0 12 M = n+1;
wolffd@0 13 Y = n+2;
wolffd@0 14 dag = zeros(N,N);
wolffd@0 15 dag([Xs M], Y)=1;
wolffd@0 16
wolffd@0 17 dnodes = M;
wolffd@0 18 ns = zeros(1, N);
wolffd@0 19 sz = 2;
wolffd@0 20 ns(Xs) = sz;
wolffd@0 21 ns(M) = n;
wolffd@0 22 ns(Y) = sz;
wolffd@0 23
wolffd@0 24 bnet = mk_bnet(dag, ns, 'discrete', M, 'observed', [M Y]);
wolffd@0 25
wolffd@0 26 psz = ns(Xs(1));
wolffd@0 27 selfsz = ns(Y);
wolffd@0 28
wolffd@0 29 W = randn(selfsz, psz);
wolffd@0 30 mu = randn(selfsz, 1);
wolffd@0 31 Sigma = eye(selfsz, selfsz);
wolffd@0 32
wolffd@0 33 bnet.CPD{M} = root_CPD(bnet, M);
wolffd@0 34 for i=Xs(:)'
wolffd@0 35 bnet.CPD{i} = gaussian_CPD(bnet, i, 'mean', zeros(psz, 1), 'cov', eye(psz, psz));
wolffd@0 36 end
wolffd@0 37 bnet.CPD{Y} = gmux_CPD(bnet, Y, 'mean', mu, 'weights', W, 'cov', Sigma);
wolffd@0 38
wolffd@0 39 evidence = cell(1,N);
wolffd@0 40 yval = randn(selfsz, 1);
wolffd@0 41 evidence{Y} = yval;
wolffd@0 42 m = 2;
wolffd@0 43 %notm = not(m-1)+1; % only valid for n=2
wolffd@0 44 notm = mysetdiff(1:n, m);
wolffd@0 45 evidence{M} = m;
wolffd@0 46
wolffd@0 47 engines = {};
wolffd@0 48 engines{end+1} = jtree_inf_engine(bnet);
wolffd@0 49 engines{end+1} = pearl_inf_engine(bnet, 'protocol', 'parallel');
wolffd@0 50
wolffd@0 51 for e=1:length(engines)
wolffd@0 52 engines{e} = enter_evidence(engines{e}, evidence);
wolffd@0 53 mXm{e} = marginal_nodes(engines{e}, Xs(m));
wolffd@0 54
wolffd@0 55 % Since M=m, only Xm was updated.
wolffd@0 56 % Hence the posterior on Xnotm should equal the prior.
wolffd@0 57 for i=notm(:)'
wolffd@0 58 mXnotm = marginal_nodes(engines{e}, Xs(i));
wolffd@0 59 assert(approxeq(mXnotm.mu, zeros(psz,1)))
wolffd@0 60 assert(approxeq(mXnotm.Sigma, eye(psz, psz)))
wolffd@0 61 end
wolffd@0 62 end
wolffd@0 63
wolffd@0 64 % Check that all engines give the same posterior
wolffd@0 65 for e=2:length(engines)
wolffd@0 66 assert(approxeq(mXm{e}.mu, mXm{1}.mu))
wolffd@0 67 assert(approxeq(mXm{e}.Sigma, mXm{1}.Sigma))
wolffd@0 68 end
wolffd@0 69
wolffd@0 70
wolffd@0 71 % Compute the correct posterior by building Xm -> Y
wolffd@0 72
wolffd@0 73 N = 2;
wolffd@0 74 dag = zeros(N,N);
wolffd@0 75 dag(1, 2)=1;
wolffd@0 76 ns = [psz selfsz];
wolffd@0 77 bnet = mk_bnet(dag, ns, 'discrete', [], 'observed', 2);
wolffd@0 78
wolffd@0 79 bnet.CPD{1} = gaussian_CPD(bnet, 1, 'mean', zeros(psz, 1), 'cov', eye(psz, psz));
wolffd@0 80 bnet.CPD{2} = gaussian_CPD(bnet, 2, 'mean', mu, 'cov', Sigma, 'weights', W);
wolffd@0 81
wolffd@0 82 jengine = jtree_inf_engine(bnet);
wolffd@0 83 evidence = {[], yval};
wolffd@0 84 jengine = enter_evidence(jengine, evidence); % apply Bayes rule to invert the arc
wolffd@0 85 mX = marginal_nodes(jengine, 1);
wolffd@0 86
wolffd@0 87 for e=1:length(engines)
wolffd@0 88 assert(approxeq(mX.mu, mXm{e}.mu))
wolffd@0 89 assert(approxeq(mX.Sigma, mXm{e}.Sigma))
wolffd@0 90 end