wolffd@0: function CPD = tabular_CPD(bnet, self, varargin)
wolffd@0: % TABULAR_CPD Make a multinomial conditional prob. distrib. (CPT)
wolffd@0: %
wolffd@0: % CPD = tabular_CPD(bnet, node) creates a random CPT.
wolffd@0: %
wolffd@0: % The following arguments can be specified [default in brackets]
wolffd@0: %
wolffd@0: % CPT - specifies the params ['rnd']
wolffd@0: %   - T means use table T; it will be reshaped to the size of node's family.
wolffd@0: %   - 'rnd' creates rnd params (drawn from uniform)
wolffd@0: %   - 'unif' creates a uniform distribution
wolffd@0: % adjustable - 0 means don't adjust the parameters during learning [1]
wolffd@0: % prior_type - defines type of prior ['none']
wolffd@0: %  - 'none' means do ML estimation
wolffd@0: %  - 'dirichlet' means add pseudo-counts to every cell
wolffd@0: %  - 'entropic' means use a prior P(theta) propto exp(-H(theta)) (see Brand)
wolffd@0: % dirichlet_weight - equivalent sample size (ess) of the dirichlet prior [1]
wolffd@0: % dirichlet_type - defines the type of Dirichlet prior ['BDeu']
wolffd@0: %  - 'unif' means put dirichlet_weight in every cell
wolffd@0: %  - 'BDeu' means we put 'dirichlet_weight/(r q)' in every cell
wolffd@0: %    where r = self_sz and q = prod(parent_sz) (see Heckerman)
wolffd@0: % trim - 1 means trim redundant params (rows in CPT) when using entropic prior [0]
wolffd@0: % entropic_pcases - list of assignments to the parents nodes when we should use 
wolffd@0: %      the entropic prior; all other cases will be estimated using ML [1:psz]
wolffd@0: % sparse - 1 means use 1D sparse array to represent CPT [0]
wolffd@0: %
wolffd@0: % e.g., tabular_CPD(bnet, i, 'CPT', T)
wolffd@0: % e.g., tabular_CPD(bnet, i, 'CPT', 'unif', 'dirichlet_weight', 2, 'dirichlet_type', 'unif')
wolffd@0: %
wolffd@0: % REFERENCES
wolffd@0: % M. Brand - "Structure learning in conditional probability models via an entropic  prior
wolffd@0: %   and parameter extinction", Neural Computation 11 (1999): 1155--1182
wolffd@0: % M. Brand - "Pattern discovery via entropy minimization" [covers annealing]
wolffd@0: %   AI & Statistics 1999. Equation numbers refer to this paper, which is available from
wolffd@0: %   www.merl.com/reports/docs/TR98-21.pdf
wolffd@0: % D. Heckerman, D. Geiger and M. Chickering, 
wolffd@0: %   "Learning Bayesian networks: the combination of knowledge and statistical data",
wolffd@0: %   Microsoft Research Tech Report, 1994
wolffd@0: 
wolffd@0: 
wolffd@0: if nargin==0
wolffd@0:   % This occurs if we are trying to load an object from a file.
wolffd@0:   CPD = init_fields;
wolffd@0:   CPD = class(CPD, 'tabular_CPD', discrete_CPD(0, []));
wolffd@0:   return;
wolffd@0: elseif isa(bnet, 'tabular_CPD')
wolffd@0:   % This might occur if we are copying an object.
wolffd@0:   CPD = bnet;
wolffd@0:   return;
wolffd@0: end
wolffd@0: CPD = init_fields;
wolffd@0: 
wolffd@0: ns = bnet.node_sizes;
wolffd@0: ps = parents(bnet.dag, self);
wolffd@0: fam_sz = ns([ps self]);
wolffd@0: psz = prod(ns(ps));
wolffd@0: CPD.sizes = fam_sz;
wolffd@0: CPD.leftright = 0;
wolffd@0: CPD.sparse = 0;
wolffd@0: 
wolffd@0: % set defaults
wolffd@0: CPD.CPT = mk_stochastic(myrand(fam_sz));
wolffd@0: CPD.adjustable = 1;
wolffd@0: CPD.prior_type = 'none';
wolffd@0: dirichlet_type = 'BDeu';
wolffd@0: dirichlet_weight = 1;
wolffd@0: CPD.trim = 0;
wolffd@0: selfprob = 0.1;
wolffd@0: CPD.entropic_pcases = 1:psz;
wolffd@0: 
wolffd@0: % extract optional args
wolffd@0: args = varargin;
wolffd@0: % check for old syntax CPD(bnet, i, CPT) as opposed to CPD(bnet, i, 'CPT', CPT)
wolffd@0: if ~isempty(args) & ~isstr(args{1})
wolffd@0:   CPD.CPT = myreshape(args{1}, fam_sz);
wolffd@0:   args = [];
wolffd@0: end
wolffd@0: 
wolffd@0: for i=1:2:length(args)
wolffd@0:   switch args{i},
wolffd@0:    case 'CPT',
wolffd@0:     T = args{i+1};
wolffd@0:     if ischar(T)
wolffd@0:       switch T
wolffd@0:        case 'unif', CPD.CPT = mk_stochastic(myones(fam_sz));
wolffd@0:        case 'rnd',  CPD.CPT = mk_stochastic(myrand(fam_sz));
wolffd@0:        otherwise,   error(['invalid CPT ' T]);       
wolffd@0:       end
wolffd@0:     else
wolffd@0:       CPD.CPT = myreshape(T, fam_sz);
wolffd@0:     end
wolffd@0:    case 'prior_type', CPD.prior_type = args{i+1};
wolffd@0:    case 'dirichlet_type', dirichlet_type = args{i+1};
wolffd@0:    case 'dirichlet_weight', dirichlet_weight = args{i+1};
wolffd@0:    case 'adjustable', CPD.adjustable = args{i+1};
wolffd@0:    case 'clamped', CPD.adjustable = ~args{i+1};
wolffd@0:    case 'trim', CPD.trim = args{i+1};
wolffd@0:    case 'entropic_pcases', CPD.entropic_pcases = args{i+1};
wolffd@0:    case 'sparse', CPD.sparse = args{i+1};
wolffd@0:    otherwise, error(['invalid argument name: ' args{i}]);       
wolffd@0:   end
wolffd@0: end
wolffd@0: 
wolffd@0: switch CPD.prior_type
wolffd@0:  case 'dirichlet',
wolffd@0:   switch dirichlet_type
wolffd@0:    case 'unif', CPD.dirichlet = dirichlet_weight * myones(fam_sz);
wolffd@0:    case 'BDeu',  CPD.dirichlet = (dirichlet_weight/psz) * mk_stochastic(myones(fam_sz));
wolffd@0:    otherwise, error(['invalid dirichlet_type ' dirichlet_type])
wolffd@0:   end
wolffd@0:  case {'entropic', 'none'}
wolffd@0:   CPD.dirichlet = [];
wolffd@0:  otherwise, error(['invalid prior_type ' prior_type])
wolffd@0: end
wolffd@0: 
wolffd@0:   
wolffd@0: 
wolffd@0: % fields to do with learning
wolffd@0: if ~CPD.adjustable
wolffd@0:   CPD.counts = [];
wolffd@0:   CPD.nparams = 0;
wolffd@0:   CPD.nsamples = [];
wolffd@0: else
wolffd@0:   %CPD.counts = zeros(size(CPD.CPT));
wolffd@0:   CPD.counts = zeros(prod(size(CPD.CPT)), 1);
wolffd@0:   psz = fam_sz(1:end-1);
wolffd@0:   ss = fam_sz(end);
wolffd@0:   if CPD.leftright
wolffd@0:     % For each of the Qps contexts, we specify Q elements on the diagoanl
wolffd@0:     CPD.nparams = Qps * Q;
wolffd@0:   else
wolffd@0:     % sum-to-1 constraint reduces the effective arity of the node by 1
wolffd@0:     CPD.nparams = prod([psz ss-1]);
wolffd@0:   end
wolffd@0:   CPD.nsamples = 0;
wolffd@0: end
wolffd@0: 
wolffd@0: CPD.trimmed_trans = [];
wolffd@0: fam_sz = CPD.sizes;
wolffd@0: 
wolffd@0: %psz = prod(fam_sz(1:end-1));
wolffd@0: %ssz = fam_sz(end);
wolffd@0: %CPD.trimmed_trans = zeros(psz, ssz); % must declare before reading
wolffd@0: 
wolffd@0: %sparse CPT
wolffd@0: if CPD.sparse
wolffd@0:    CPD.CPT = sparse(CPD.CPT(:));
wolffd@0: end
wolffd@0: 
wolffd@0: CPD = class(CPD, 'tabular_CPD', discrete_CPD(~CPD.adjustable, fam_sz));
wolffd@0: 
wolffd@0: 
wolffd@0: %%%%%%%%%%%
wolffd@0: 
wolffd@0: function CPD = init_fields()
wolffd@0: % This ensures we define the fields in the same order 
wolffd@0: % no matter whether we load an object from a file,
wolffd@0: % or create it from scratch. (Matlab requires this.)
wolffd@0: 
wolffd@0: CPD.CPT = [];
wolffd@0: CPD.sizes = [];
wolffd@0: CPD.prior_type = [];
wolffd@0: CPD.dirichlet = [];
wolffd@0: CPD.adjustable = [];
wolffd@0: CPD.counts = [];
wolffd@0: CPD.nparams = [];
wolffd@0: CPD.nsamples = [];
wolffd@0: CPD.trim = [];
wolffd@0: CPD.trimmed_trans = [];
wolffd@0: CPD.leftright = [];
wolffd@0: CPD.entropic_pcases = [];
wolffd@0: CPD.sparse = [];
wolffd@0: