wolffd@0: function CPD = gaussian_CPD(varargin)
wolffd@0: % GAUSSIAN_CPD Make a conditional linear Gaussian distrib.
wolffd@0: %
wolffd@0: % To define this CPD precisely, call the continuous (cts) parents (if any) X,
wolffd@0: % the discrete parents (if any) Q, and this node Y. Then the distribution on Y is:
wolffd@0: % - no parents: Y ~ N(mu, Sigma)
wolffd@0: % - cts parents : Y|X=x ~ N(mu + W x, Sigma)
wolffd@0: % - discrete parents: Y|Q=i ~ N(mu(i), Sigma(i))
wolffd@0: % - cts and discrete parents: Y|X=x,Q=i ~ N(mu(i) + W(i) x, Sigma(i))
wolffd@0: %
wolffd@0: % CPD = gaussian_CPD(bnet, node, ...) will create a CPD with random parameters,
wolffd@0: % where node is the number of a node in this equivalence class.
wolffd@0: %
wolffd@0: % The list below gives optional arguments [default value in brackets].
wolffd@0: % (Let ns(i) be the size of node i, X = ns(X), Y = ns(Y) and Q = prod(ns(Q)).)
wolffd@0: %
wolffd@0: % mean       - mu(:,i) is the mean given Q=i [ randn(Y,Q) ]
wolffd@0: % cov        - Sigma(:,:,i) is the covariance given Q=i [ repmat(eye(Y,Y), [1 1 Q]) ]
wolffd@0: % weights    - W(:,:,i) is the regression matrix given Q=i [ randn(Y,X,Q) ]
wolffd@0: % cov_type   - if 'diag', Sigma(:,:,i) is diagonal [ 'full' ]
wolffd@0: % tied_cov   - if 1, we constrain Sigma(:,:,i) to be the same for all i [0]
wolffd@0: % clamp_mean - if 1, we do not adjust mu(:,i) during learning [0]
wolffd@0: % clamp_cov  - if 1, we do not adjust Sigma(:,:,i) during learning [0]
wolffd@0: % clamp_weights - if 1, we do not adjust W(:,:,i) during learning [0]
wolffd@0: % cov_prior_weight - weight given to I prior for estimating Sigma [0.01]
wolffd@0: %
wolffd@0: % e.g., CPD = gaussian_CPD(bnet, i, 'mean', [0; 0], 'clamp_mean', 'yes')
wolffd@0: %
wolffd@0: % For backwards compatibility with BNT2, you can also specify the parameters in the following order
wolffd@0: %   CPD = gaussian_CPD(bnet, self, mu, Sigma, W, cov_type, tied_cov, clamp_mean, clamp_cov, clamp_weight)
wolffd@0: %
wolffd@0: % Sometimes it is useful to create an "isolated" CPD, without needing to pass in a bnet.
wolffd@0: % In this case, you must specify the discrete and cts parents (dps, cps) and the family sizes, followed
wolffd@0: % by the optional arguments above:
wolffd@0: %   CPD = gaussian_CPD('self', i, 'dps', dps, 'cps', cps, 'sz', fam_size, ...)
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:   clamp = 0;
wolffd@0:   CPD = class(CPD, 'gaussian_CPD', generic_CPD(clamp));
wolffd@0:   return;
wolffd@0: elseif isa(varargin{1}, 'gaussian_CPD')
wolffd@0:   % This might occur if we are copying an object.
wolffd@0:   CPD = varargin{1};
wolffd@0:   return;
wolffd@0: end
wolffd@0: CPD = init_fields;
wolffd@0:  
wolffd@0: CPD = class(CPD, 'gaussian_CPD', generic_CPD(0));
wolffd@0: 
wolffd@0: 
wolffd@0: % parse mandatory arguments
wolffd@0: if ~isstr(varargin{1}) % pass in bnet
wolffd@0:   bnet = varargin{1};
wolffd@0:   self = varargin{2};
wolffd@0:   args = varargin(3:end);
wolffd@0:   ns = bnet.node_sizes;
wolffd@0:   ps = parents(bnet.dag, self);
wolffd@0:   dps = myintersect(ps, bnet.dnodes);
wolffd@0:   cps = myintersect(ps, bnet.cnodes);
wolffd@0:   fam_sz = ns([ps self]);
wolffd@0: else
wolffd@0:   disp('parsing new style')
wolffd@0:   for i=1:2:length(varargin)
wolffd@0:     switch varargin{i},
wolffd@0:      case 'self', self = varargin{i+1}; 
wolffd@0:      case 'dps',  dps = varargin{i+1};
wolffd@0:      case 'cps',  cps = varargin{i+1};
wolffd@0:      case 'sz',   fam_sz = varargin{i+1};
wolffd@0:     end
wolffd@0:   end
wolffd@0:   ps = myunion(dps, cps);
wolffd@0:   args = varargin;
wolffd@0: end
wolffd@0: 
wolffd@0: CPD.self = self;
wolffd@0: CPD.sizes = fam_sz;
wolffd@0: 
wolffd@0: % Figure out which (if any) of the parents are discrete, and which cts, and how big they are
wolffd@0: % dps = discrete parents, cps = cts parents
wolffd@0: CPD.cps = find_equiv_posns(cps, ps); % cts parent index
wolffd@0: CPD.dps = find_equiv_posns(dps, ps);
wolffd@0: ss = fam_sz(end);
wolffd@0: psz = fam_sz(1:end-1);
wolffd@0: dpsz = prod(psz(CPD.dps));
wolffd@0: cpsz = sum(psz(CPD.cps));
wolffd@0: 
wolffd@0: % set default params
wolffd@0: CPD.mean = randn(ss, dpsz);
wolffd@0: CPD.cov = 100*repmat(eye(ss), [1 1 dpsz]);    
wolffd@0: CPD.weights = randn(ss, cpsz, dpsz);
wolffd@0: CPD.cov_type = 'full';
wolffd@0: CPD.tied_cov = 0;
wolffd@0: CPD.clamped_mean = 0;
wolffd@0: CPD.clamped_cov = 0;
wolffd@0: CPD.clamped_weights = 0;
wolffd@0: CPD.cov_prior_weight = 0.01;
wolffd@0: 
wolffd@0: nargs = length(args);
wolffd@0: if nargs > 0
wolffd@0:   if ~isstr(args{1})
wolffd@0:     % gaussian_CPD(bnet, self, mu, Sigma, W, cov_type, tied_cov, clamp_mean, clamp_cov, clamp_weights)
wolffd@0:     if nargs >= 1 & ~isempty(args{1}), CPD.mean = args{1}; end
wolffd@0:     if nargs >= 2 & ~isempty(args{2}), CPD.cov = args{2}; end
wolffd@0:     if nargs >= 3 & ~isempty(args{3}), CPD.weights = args{3}; end
wolffd@0:     if nargs >= 4 & ~isempty(args{4}), CPD.cov_type = args{4}; end
wolffd@0:     if nargs >= 5 & ~isempty(args{5}) & strcmp(args{5}, 'tied'), CPD.tied_cov = 1; end
wolffd@0:     if nargs >= 6 & ~isempty(args{6}), CPD.clamped_mean = 1; end
wolffd@0:     if nargs >= 7 & ~isempty(args{7}), CPD.clamped_cov = 1; end
wolffd@0:     if nargs >= 8 & ~isempty(args{8}), CPD.clamped_weights = 1; end
wolffd@0:   else
wolffd@0:     CPD = set_fields(CPD, args{:});
wolffd@0:   end
wolffd@0: end
wolffd@0: 
wolffd@0: % Make sure the matrices have 1 dimension per discrete parent.
wolffd@0: % Bug fix due to Xuejing Sun 3/6/01
wolffd@0: CPD.mean = myreshape(CPD.mean, [ss ns(dps)]);
wolffd@0: CPD.cov = myreshape(CPD.cov, [ss ss ns(dps)]);
wolffd@0: CPD.weights = myreshape(CPD.weights, [ss cpsz ns(dps)]);
wolffd@0:   
wolffd@0: CPD.init_cov = CPD.cov;  % we reset to this if things go wrong during learning
wolffd@0: 
wolffd@0: % expected sufficient statistics 
wolffd@0: CPD.Wsum = zeros(dpsz,1);
wolffd@0: CPD.WYsum = zeros(ss, dpsz);
wolffd@0: CPD.WXsum = zeros(cpsz, dpsz);
wolffd@0: CPD.WYYsum = zeros(ss, ss, dpsz);
wolffd@0: CPD.WXXsum = zeros(cpsz, cpsz, dpsz);
wolffd@0: CPD.WXYsum = zeros(cpsz, ss, dpsz);
wolffd@0: 
wolffd@0: % For BIC
wolffd@0: CPD.nsamples = 0;
wolffd@0: switch CPD.cov_type
wolffd@0:   case 'full',
wolffd@0:     ncov_params = ss*(ss-1)/2; % since symmetric (and positive definite)
wolffd@0:   case 'diag',
wolffd@0:     ncov_params = ss;
wolffd@0:   otherwise
wolffd@0:     error(['unrecognized cov_type ' cov_type]);
wolffd@0: end
wolffd@0: % params = weights + mean + cov
wolffd@0: if CPD.tied_cov
wolffd@0:   CPD.nparams = ss*cpsz*dpsz + ss*dpsz + ncov_params;
wolffd@0: else
wolffd@0:   CPD.nparams = ss*cpsz*dpsz + ss*dpsz + dpsz*ncov_params;
wolffd@0: end
wolffd@0: 
wolffd@0: 
wolffd@0: 
wolffd@0: clamped = CPD.clamped_mean & CPD.clamped_cov & CPD.clamped_weights;
wolffd@0: CPD = set_clamped(CPD, clamped);
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.self = [];
wolffd@0: CPD.sizes = [];
wolffd@0: CPD.cps = [];
wolffd@0: CPD.dps = [];
wolffd@0: CPD.mean = [];
wolffd@0: CPD.cov = [];
wolffd@0: CPD.weights = [];
wolffd@0: CPD.clamped_mean = [];
wolffd@0: CPD.clamped_cov = [];
wolffd@0: CPD.clamped_weights = [];
wolffd@0: CPD.init_cov = [];
wolffd@0: CPD.cov_type = [];
wolffd@0: CPD.tied_cov = [];
wolffd@0: CPD.Wsum = [];
wolffd@0: CPD.WYsum = [];
wolffd@0: CPD.WXsum = [];
wolffd@0: CPD.WYYsum = [];
wolffd@0: CPD.WXXsum = [];
wolffd@0: CPD.WXYsum = [];
wolffd@0: CPD.nsamples = [];
wolffd@0: CPD.nparams = [];            
wolffd@0: CPD.cov_prior_weight = [];