www.pudn.com > ConstrainedEM.zip > reduce_local_net2.m

function  engine=reduce_local_net2( engine, pots); 
% reduce a connected inference net to a tree by weighting the arcs and calling 
% minimal spanning tree algorithm 
% the arc weights are determined by the potentials at both ends of the arc .  it 
% is p1*p2'. 
% engine - a struct with mnet containing the adjacency matrix. 
% pots - (number of nodes * variables number of values) table containing the 
% potentials at the nodes of the graph 
 
% create a node_arc incidence matrix and costs vector for mst function 
global anti_chunks_pruned; 
global anti_chunk_num; 
adj=engine.mnet; 
k=length(adj);		% number of nodes 
[x y]=find(triu(adj)==1); 
l=length(x);		% number of arcs; 
 
if ~(k-1==l)        % if component isn't a tree already 
    pots=pots./(sum(pots')'*ones(1,size(pots,2)));	% normalize pots before determining costs  
    costs=-sum((pots(x,:).*pots(y,:))'); 
     
    % find minimal spanning tree and adjust the inference engine 
    [jnk,ind]=sort(costs); 
    mst=mstree([x(ind),y(ind),jnk'],k); 
    anti_chunks_pruned=anti_chunks_pruned+l-k+1; 
    anti_chunk_num=anti_chunk_num-(l-k+1); 
    adj=zeros(k,k); 
    adj(sub2ind([k k],mst(:,1),mst(:,2)))=1; 
    adj(sub2ind([k k],mst(:,2),mst(:,1)))=1; 
end 
 
engine = inf_engine_from_mnet(adj,k ,1);