wolffd@0: function [mpe, ll] = calc_mpe_dbn(engine, evidence, break_ties)
wolffd@0: % CALC_MPE Computes the most probable explanation of the evidence
wolffd@0: % [mpe, ll] = calc_mpe_dbn(engine, evidence, break_ties)
wolffd@0: %
wolffd@0: % INPUT
wolffd@0: % engine must support max-propagation
wolffd@0: % evidence{i,t} is the observed value of node i in slice t, or [] if hidden
wolffd@0: %
wolffd@0: % OUTPUT
wolffd@0: % mpe{i,t} is the most likely value of node i (cell array!)
wolffd@0: % ll is the log-likelihood of the globally best assignment
wolffd@0: %
wolffd@0: % This currently only works when all hidden nodes are discrete
wolffd@0: 
wolffd@0: if nargin < 3, break_ties = 0; end
wolffd@0: 
wolffd@0: if break_ties
wolffd@0:   disp('warning: break ties is ignored')
wolffd@0: end
wolffd@0: 
wolffd@0: [engine, ll] = enter_evidence(engine, evidence, 'maximize', 1);
wolffd@0: 
wolffd@0: observed = ~isemptycell(evidence);
wolffd@0: [ss T] = size(evidence);
wolffd@0: scalar = 1;
wolffd@0: N = length(evidence);
wolffd@0: mpe = cell(ss,T);
wolffd@0: bnet = bnet_from_engine(engine);
wolffd@0: ns = bnet.node_sizes;
wolffd@0: for t=1:T
wolffd@0:   for i=1:ss
wolffd@0:     m = marginal_nodes(engine, i, t);
wolffd@0:     % observed nodes are all set to 1 inside the inference engine, so we must undo this
wolffd@0:     if observed(i,t)
wolffd@0:       mpe{i,t} = evidence{i,t};
wolffd@0:     else
wolffd@0:       assert(length(m.T) == ns(i));
wolffd@0:       mpe{i,t} = argmax(m.T);
wolffd@0:     end
wolffd@0:   end
wolffd@0: end