wolffd@0: function [B,D,mu] = extract_params_from_gbn(bnet)
wolffd@0: % Extract all the local parameters of each Gaussian node, and collect them into global matrices.
wolffd@0: % [B,D,mu] = extract_params_from_gbn(bnet)
wolffd@0: %
wolffd@0: % B(i,j) is a block matrix that contains the transposed weight matrix from node i to node j.
wolffd@0: % D(i,i) is a block matrix that contains the noise covariance matrix for node i.
wolffd@0: % mu(i) is a block vector that contains the shifted noise mean for node i.
wolffd@0: 
wolffd@0: % In Shachter's model, the mean of each node in the global gaussian is
wolffd@0: % the same as the node's local unconditional mean.
wolffd@0: % In Alag's model (which we use), the global mean gets shifted.
wolffd@0: 
wolffd@0: 
wolffd@0: num_nodes = length(bnet.dag);
wolffd@0: bs = bnet.node_sizes(:); % bs = block sizes
wolffd@0: N = sum(bs); % num scalar nodes
wolffd@0: 
wolffd@0: B = zeros(N,N);
wolffd@0: D = zeros(N,N);
wolffd@0: mu = zeros(N,1);
wolffd@0: 
wolffd@0: for i=1:num_nodes % in topological order
wolffd@0:   ps = parents(bnet.dag, i);
wolffd@0:   e = bnet.equiv_class(i);
wolffd@0:   %[m, Sigma, weights] = extract_params_from_CPD(bnet.CPD{e});
wolffd@0:   s = struct(bnet.CPD{e}); % violate privacy of object
wolffd@0:   m = s.mean; Sigma = s.cov; weights = s.weights;
wolffd@0:   if length(ps) == 0
wolffd@0:     mu(block(i,bs)) = m;
wolffd@0:   else
wolffd@0:     mu(block(i,bs)) = m + weights *  mu(block(ps,bs));
wolffd@0:   end
wolffd@0:   B(block(ps,bs), block(i,bs)) = weights';
wolffd@0:   D(block(i,bs), block(i,bs)) = Sigma;
wolffd@0: end
wolffd@0: 
wolffd@0: 
wolffd@0: