Daniel@0: function mpe = find_mpe(engine, evidence, varargin)
Daniel@0: % FIND_MPE Find the most probable explanation of the data  (belprop)
Daniel@0: % function mpe = find_mpe(engine, evidence,...)
Daniel@0: %
Daniel@0: % evidence{i} = [] if X(i) is hidden, and otherwise contains its observed value (scalar or column vector).
Daniel@0: %
Daniel@0: % This finds the marginally most likely value for each hidden node,
Daniel@0: % and may give the wrong results even if the graph is acyclic,
Daniel@0: % unless you set break_ties = 1.
Daniel@0: %
Daniel@0: % The following optional arguments can be specified in the form of name/value pairs:
Daniel@0: % [default value in brackets]
Daniel@0: %
Daniel@0: % break_ties is optional. If 1, we will force ties to be broken consistently
Daniel@0: %  by calling enter_evidence N times. (see Jensen96, p106) Default = 1.
Daniel@0: 
Daniel@0: break_ties = 1;
Daniel@0: 
Daniel@0: % parse optional params
Daniel@0: args = varargin;
Daniel@0: nargs = length(args);
Daniel@0: for i=1:2:nargs
Daniel@0:   switch args{i},
Daniel@0:    case 'break_ties',    break_ties = args{i+1}; 
Daniel@0:    otherwise,  
Daniel@0:     error(['invalid argument name ' args{i}]);       
Daniel@0:   end
Daniel@0: end
Daniel@0: 
Daniel@0: engine = enter_evidence(engine, evidence, 'maximize', 1);
Daniel@0: 
Daniel@0: observed = ~isemptycell(evidence);
Daniel@0: evidence = evidence(:); % hack to handle unrolled DBNs
Daniel@0: N = length(evidence);
Daniel@0: mpe = cell(1,N);
Daniel@0: for i=1:N
Daniel@0:   m = marginal_nodes(engine, i);
Daniel@0:   % observed nodes are all set to 1 inside the inference engine, so we must undo this
Daniel@0:   if observed(i)
Daniel@0:     mpe{i} = evidence{i};
Daniel@0:   else
Daniel@0:     mpe{i} = argmax(m.T);
Daniel@0:     if break_ties
Daniel@0:       evidence{i} = mpe{i};                             
Daniel@0:       [engine, ll] = enter_evidence(engine, evidence, 'maximize', 1);  
Daniel@0:     end
Daniel@0:   end
Daniel@0: end
Daniel@0: