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