wolffd@0: function pot = convert_to_pot(CPD, pot_type, domain, evidence)
wolffd@0: % CONVERT_TO_POT Convert a gmux CPD to a Gaussian potential
wolffd@0: % pot = convert_to_pot(CPD, pot_type, domain, evidence)
wolffd@0:   
wolffd@0: switch pot_type
wolffd@0:  case {'d', 'u', 'cg', 'scg'},
wolffd@0:   error(['can''t convert gmux to potential of type ' pot_type])
wolffd@0: 
wolffd@0:  case {'c','g'},
wolffd@0:   % We create a large weight matrix with zeros in all blocks corresponding
wolffd@0:   % to the non-chosen parents, since they are effectively disconnected.
wolffd@0:   % The chosen parent is determined by the value, m,  of the discrete parent.
wolffd@0:   % Thus the potential is as large as the whole family.
wolffd@0:   ps = domain(1:end-1);
wolffd@0:   dps = ps(CPD.dps); % CPD.dps is an index, not a node number (because of param tying)
wolffd@0:   cps = ps(CPD.cps);
wolffd@0:   m = evidence{dps};
wolffd@0:   if isempty(m)
wolffd@0:     error('gmux node must have observed discrete parent')
wolffd@0:   end
wolffd@0:   bs = CPD.sizes(CPD.cps);
wolffd@0:   b = block(m, bs);
wolffd@0:   sum_cpsz = sum(CPD.sizes(CPD.cps));
wolffd@0:   selfsz = CPD.sizes(end);
wolffd@0:   W = zeros(selfsz, sum_cpsz);
wolffd@0:   W(:,b) = CPD.weights(:,:,m);
wolffd@0: 
wolffd@0:   ns = zeros(1, max(domain));
wolffd@0:   ns(domain) = CPD.sizes;
wolffd@0:   self = domain(end);
wolffd@0:   cdom = [cps(:)' self];
wolffd@0:   pot = linear_gaussian_to_cpot(CPD.mean(:,m), CPD.cov(:,:,m), W, domain, ns, cdom, evidence);
wolffd@0:   
wolffd@0:  otherwise,
wolffd@0:   error(['unrecognized pot_type' pot_type])
wolffd@0: end
wolffd@0: