wolffd@0: function CPD = hhmmQ_CPD(bnet, self, varargin)
wolffd@0: % HHMMQ_CPD Make the CPD for a Q node in a hierarchical HMM
wolffd@0: % CPD = hhmmQ_CPD(bnet, self, ...)
wolffd@0: %
wolffd@0: %  Fself(t-1)   Qps(t)
wolffd@0: %           \    |
wolffd@0: %            \   v
wolffd@0: %  Qold(t-1) ->  Q(t)
wolffd@0: %            /
wolffd@0: %           /
wolffd@0: %  Fbelow(t-1) 
wolffd@0: %
wolffd@0: % Let ss = slice size = num. nodes per slice.
wolffd@0: % This node is Q(t), and has mandatory parents Qold(t-1) (assumed to be numbered Q(t)-ss)
wolffd@0: % and optional parents Fbelow, Fself, Qps.
wolffd@0: % We require parents to be ordered (numbered) as follows:
wolffd@0: % Qold, Fbelow, Fself, Qps, Q.
wolffd@0: %
wolffd@0: % If Fself=2, we use the transition matrix, else we use the prior matrix.
wolffd@0: % If Fself node is omitted (eg. top level), we always use the transition matrix.
wolffd@0: % If Fbelow=2, we may change state, otherwise we must stay in the same state.
wolffd@0: % If Fbelow node is omitted (eg., bottom level), we may change state at every step.
wolffd@0: % If Qps (Q parents) are specified, all parameters are conditioned on their joint value.
wolffd@0: % We may choose any subset of nodes to condition on, as long as they as numbered lower than self.
wolffd@0: %
wolffd@0: % optional args [defaults]
wolffd@0: %
wolffd@0: % Fself - node number <= ss
wolffd@0: % Fbelow  - node number  <= ss
wolffd@0: % Qps - node numbers (all <= 2*ss) - uses 2TBN indexing
wolffd@0: % transprob - transprob(i,k,j) = prob transition from i to j given Qps = k ['leftright']
wolffd@0: % selfprob  - prob of a transition from i to i given Qps=k [0.1]
wolffd@0: % startprob - startprob(k,j) = prob start in j given Qps = k ['leftstart']
wolffd@0: % startargs - other args to be passed to the sub tabular_CPD for learning startprob
wolffd@0: % transargs - other args will be passed to the sub tabular_CPD for learning transprob
wolffd@0: % fullstartprob - 1 means startprob depends on Q(t-1) [0]
wolffd@0: % hhmmQ_CPD is a subclass of tabular_CPD so we inherit inference methods like CPD_to_pot, etc.
wolffd@0: %
wolffd@0: % We create isolated tabular_CPDs with no F parents to learn transprob/startprob
wolffd@0: % so we can avail of e.g., entropic or Dirichlet priors.
wolffd@0: % In the future, we will be able to represent the transprob using a tree_CPD.
wolffd@0: %
wolffd@0: % For details, see "Linear-time inference in hierarchical HMMs", Murphy and Paskin, NIPS'01.
wolffd@0: 
wolffd@0: 
wolffd@0: ss = bnet.nnodes_per_slice;
wolffd@0: ns = bnet.node_sizes(:);
wolffd@0: 
wolffd@0: % set default arguments
wolffd@0: Fself = [];
wolffd@0: Fbelow = [];
wolffd@0: Qps = [];
wolffd@0: startprob = 'leftstart';
wolffd@0: transprob = 'leftright';
wolffd@0: startargs = {};
wolffd@0: transargs = {};
wolffd@0: selfprob = 0.1;
wolffd@0: fullstartprob = 0;
wolffd@0: 
wolffd@0: for i=1:2:length(varargin)
wolffd@0:   switch varargin{i},
wolffd@0:    case 'Fself', Fself = varargin{i+1};
wolffd@0:    case 'Fbelow', Fbelow = varargin{i+1};
wolffd@0:    case 'Qps', Qps = varargin{i+1};
wolffd@0:    case 'transprob', transprob = varargin{i+1}; 
wolffd@0:    case 'selfprob',  selfprob = varargin{i+1}; 
wolffd@0:    case 'startprob', startprob = varargin{i+1}; 
wolffd@0:    case 'startargs', startargs = varargin{i+1}; 
wolffd@0:    case 'transargs', transargs = varargin{i+1}; 
wolffd@0:    case 'fullstartprob', fullstartprob = varargin{i+1}; 
wolffd@0:   end
wolffd@0: end
wolffd@0: 
wolffd@0: CPD.fullstartprob = fullstartprob;
wolffd@0: 
wolffd@0: ps = parents(bnet.dag, self);
wolffd@0: ndsz = ns(:)';
wolffd@0: CPD.dom_sz = [ndsz(ps) ns(self)];
wolffd@0: CPD.Fself_ndx = find_equiv_posns(Fself, ps);
wolffd@0: CPD.Fbelow_ndx = find_equiv_posns(Fbelow, ps);
wolffd@0: %CPD.Qps_ndx = find_equiv_posns(Qps+ss, ps);
wolffd@0: CPD.Qps_ndx = find_equiv_posns(Qps, ps);
wolffd@0: old_self = self-ss;
wolffd@0: CPD.old_self_ndx = find_equiv_posns(old_self, ps);
wolffd@0: 
wolffd@0: Qps = ps(CPD.Qps_ndx);
wolffd@0: CPD.Qsz = ns(self);
wolffd@0: CPD.Qpsz = prod(ns(Qps));
wolffd@0: CPD.Qpsizes = ns(Qps);
wolffd@0: Qsz = CPD.Qsz;
wolffd@0: Qpsz = CPD.Qpsz;
wolffd@0: 
wolffd@0: if strcmp(transprob, 'leftright')
wolffd@0:   LR = mk_leftright_transmat(Qsz, selfprob);
wolffd@0:   transprob = repmat(reshape(LR, [1 Qsz Qsz]), [Qpsz 1 1]); % transprob(k,i,j)
wolffd@0:   transprob = permute(transprob, [2 1 3]); % now transprob(i,k,j)
wolffd@0: end
wolffd@0: transargs{end+1} = 'CPT';
wolffd@0: transargs{end+1} = transprob;
wolffd@0: CPD.sub_CPD_trans = mk_isolated_tabular_CPD(ns([old_self Qps self]), transargs);
wolffd@0: S = struct(CPD.sub_CPD_trans);
wolffd@0: %CPD.transprob = myreshape(S.CPT, [Qsz Qpsz Qsz]);
wolffd@0: CPD.transprob = S.CPT;
wolffd@0: 
wolffd@0: 
wolffd@0: if strcmp(startprob, 'leftstart')
wolffd@0:   startprob = zeros(Qpsz, Qsz);
wolffd@0:   startprob(:,1) = 1;
wolffd@0: end
wolffd@0: if isempty(CPD.Fself_ndx)
wolffd@0:   CPD.sub_CPD_start = [];
wolffd@0:   CPD.startprob = [];
wolffd@0: else
wolffd@0:   startargs{end+1} = 'CPT';
wolffd@0:   startargs{end+1} = startprob;
wolffd@0:   if CPD.fullstartprob
wolffd@0:     CPD.sub_CPD_start = mk_isolated_tabular_CPD(ns([self Qps self]), startargs);
wolffd@0:     S = struct(CPD.sub_CPD_start);
wolffd@0:     %CPD.startprob = myreshape(S.CPT, [Qsz Qpsz Qsz]);
wolffd@0:     CPD.startprob = S.CPT;
wolffd@0:   else
wolffd@0:     CPD.sub_CPD_start = mk_isolated_tabular_CPD(ns([Qps self]), startargs);
wolffd@0:     S = struct(CPD.sub_CPD_start);
wolffd@0:     %CPD.startprob = myreshape(S.CPT, [CPD.Qpsizes Qsz]);
wolffd@0:     CPD.startprob = S.CPT;
wolffd@0:   end
wolffd@0: end
wolffd@0: 
wolffd@0: CPD = class(CPD, 'hhmmQ_CPD', tabular_CPD(bnet, self));
wolffd@0: 
wolffd@0: CPD = update_CPT(CPD);
wolffd@0: