wolffd@0: function bnet = mk_rnd_map_hhmm(varargin)
wolffd@0: 
wolffd@0: % We copy the deterministic structure of the real HHMM,
wolffd@0: % but randomize the probabilities of the adjustable CPDs.
wolffd@0: % The key trick is that 0s in the real HHMM remain 0
wolffd@0: % even when multiplied by a randon number.
wolffd@0: 
wolffd@0: obs_model = 'unique';
wolffd@0: 
wolffd@0: for i=1:2:length(varargin)
wolffd@0:   switch varargin{i},
wolffd@0:    case 'obs_model', obs_model = varargin{i+1};
wolffd@0:   end
wolffd@0: end
wolffd@0: 
wolffd@0: 
wolffd@0: unique_obs = strcmp(obs_model, 'unique');
wolffd@0: 
wolffd@0: psuccess = 0.9;
wolffd@0: % must be less than 1, so that pfail > 0
wolffd@0: % otherwise we copy too many 0s
wolffd@0: bnet = mk_map_hhmm('p', psuccess, 'obs_model', obs_model);
wolffd@0: ns = bnet.node_sizes;
wolffd@0: ss = bnet.nnodes_per_slice;
wolffd@0: 
wolffd@0: U = 1; A = 2; C = 3; F = 4;
wolffd@0: %unique_obs = (bnet.nnodes_per_slice == 5);
wolffd@0: if unique_obs
wolffd@0:   onodes = 5;
wolffd@0: else
wolffd@0:   north = 5; east = 6; south = 7; west = 8;
wolffd@0:   onodes = [north east south west];
wolffd@0: end
wolffd@0: 
wolffd@0: eclass = bnet.equiv_class;
wolffd@0: S=struct(bnet.CPD{eclass(F,1)});
wolffd@0: CPT = mk_stochastic(rand(size(S.CPT)) .* S.CPT);
wolffd@0: bnet.CPD{eclass(F,1)} = tabular_CPD(bnet, F, 'CPT', CPT);
wolffd@0: 
wolffd@0: 
wolffd@0: % Observation model
wolffd@0: if unique_obs
wolffd@0:   CPT = zeros(ns(A), ns(C), 5);
wolffd@0:   CPT(1,1,1)=1;  % Theo state 4
wolffd@0:   CPT(1,2,2)=1;  % Theo state 5
wolffd@0:   CPT(1,3,3)=1; % Theo state 6
wolffd@0:   CPT(2,1,4)=1; % Theo state 9
wolffd@0:   CPT(2,2,5)=1; % Theo state 10
wolffd@0:   %CPT(2,3,:) undefined
wolffd@0:   O = onodes(1);
wolffd@0:   bnet.CPD{eclass(O,1)} = tabular_CPD(bnet, O, 'CPT', CPT);
wolffd@0: else
wolffd@0:   for i=[north east south west]
wolffd@0:     CPT = mk_stochastic(rand(ns(A), ns(C), 2));
wolffd@0:     bnet.CPD{eclass(i,1)} = tabular_CPD(bnet, i, 'CPT', CPT);
wolffd@0:   end
wolffd@0: end
wolffd@0: 
wolffd@0: % Define the CPDs for slice 2
wolffd@0: 
wolffd@0: startprob = zeros(ns(U), ns(A));
wolffd@0: S = struct(bnet.CPD{eclass(A,2)});
wolffd@0: transprob = mk_stochastic(rand(size(S.transprob)) .* S.transprob);
wolffd@0: bnet.CPD{eclass(A,2)} = hhmm2Q_CPD(bnet, A+ss, 'Fbelow', F, ...
wolffd@0: 				  'startprob', startprob, 'transprob', transprob);
wolffd@0: 
wolffd@0: S = struct(bnet.CPD{eclass(C,2)});
wolffd@0: transprob = mk_stochastic(rand(size(S.transprob)) .* S.transprob);
wolffd@0: startprob = mk_stochastic(rand(size(S.startprob)) .* S.startprob);
wolffd@0: bnet.CPD{eclass(C,2)} = hhmm2Q_CPD(bnet, C+ss, 'Fself', F, ...
wolffd@0: 				  'startprob', startprob, 'transprob', transprob);
wolffd@0: 
wolffd@0: