wolffd@0: function dag = learn_struct_K2(data, ns, order, varargin)
wolffd@0: % LEARN_STRUCT_K2 Greedily learn the best structure compatible with a fixed node ordering
wolffd@0: % best_dag = learn_struct_K2(data, node_sizes, order, ...)
wolffd@0: %
wolffd@0: % data(i,m) = value of node i in case m (can be a cell array).
wolffd@0: % node_sizes(i) is the size of node i.
wolffd@0: % order(i) is the i'th node in the topological ordering.
wolffd@0: %
wolffd@0: % The following optional arguments can be specified in the form of name/value pairs:
wolffd@0: % [default value in brackets]
wolffd@0: %
wolffd@0: % max_fan_in - this the largest number of parents we allow per node [N]
wolffd@0: % scoring_fn - 'bayesian' or 'bic' [ 'bayesian' ]
wolffd@0: %              Currently, only networks with all tabular nodes support Bayesian scoring.
wolffd@0: % type       - type{i} is the type of CPD to use for node i, where the type is a string
wolffd@0: %              of the form 'tabular', 'noisy_or', 'gaussian', etc. [ all cells contain 'tabular' ]
wolffd@0: % params     - params{i} contains optional arguments passed to the CPD constructor for node i,
wolffd@0: %              or [] if none.  [ all cells contain {'prior', 1}, meaning use uniform Dirichlet priors ]
wolffd@0: % discrete   - the list of discrete nodes [ 1:N ]
wolffd@0: % clamped    - clamped(i,m) = 1 if node i is clamped in case m [ zeros(N, ncases) ]
wolffd@0: % verbose    - 'yes' means display output while running [ 'no' ]
wolffd@0: %
wolffd@0: % e.g., dag = learn_struct_K2(data, ns, order, 'scoring_fn', 'bic', 'params', [])
wolffd@0: %
wolffd@0: % To be backwards compatible with BNT2, you can also specify arguments as follows
wolffd@0: %   dag = learn_struct_K2(data, node_sizes, order, max_fan_in)    
wolffd@0: %
wolffd@0: % This algorithm is described in
wolffd@0: % - Cooper and Herskovits,  "A Bayesian method for the induction of probabilistic
wolffd@0: %      networks from data", Machine Learning Journal 9:308--347, 1992
wolffd@0: 
wolffd@0: [n ncases] = size(data);
wolffd@0: 
wolffd@0: % set default params
wolffd@0: type = cell(1,n);
wolffd@0: params = cell(1,n);
wolffd@0: for i=1:n
wolffd@0:   type{i} = 'tabular';
wolffd@0:   %params{i} = { 'prior', 1 };
wolffd@0:   params{i} = { 'prior_type', 'dirichlet', 'dirichlet_weight', 1 };
wolffd@0: end
wolffd@0: scoring_fn = 'bayesian';
wolffd@0: discrete = 1:n;
wolffd@0: clamped = zeros(n, ncases);
wolffd@0: 
wolffd@0: max_fan_in = n;
wolffd@0: verbose = 0;
wolffd@0: 
wolffd@0: args = varargin;
wolffd@0: nargs = length(args);
wolffd@0: if length(args) > 0 
wolffd@0:   if isstr(args{1})
wolffd@0:     for i=1:2:nargs
wolffd@0:       switch args{i},
wolffd@0:        case 'verbose',    verbose = strcmp(args{i+1}, 'yes');
wolffd@0:        case 'max_fan_in', max_fan_in = args{i+1}; 
wolffd@0:        case 'scoring_fn', scoring_fn = args{i+1};
wolffd@0:        case 'type',       type = args{i+1}; 
wolffd@0:        case 'discrete',   discrete = args{i+1}; 
wolffd@0:        case 'clamped',    clamped = args{i+1}; 
wolffd@0:        case 'params',     if isempty(args{i+1}), params = cell(1,n); else params = args{i+1};  end
wolffd@0:       end
wolffd@0:     end
wolffd@0:   else
wolffd@0:     max_fan_in = args{1};
wolffd@0:   end
wolffd@0: end
wolffd@0: 
wolffd@0: dag = zeros(n,n);
wolffd@0: 
wolffd@0: for i=1:n
wolffd@0:   ps = [];
wolffd@0:   j = order(i);
wolffd@0:   u = find(clamped(j,:)==0);    
wolffd@0:   score = score_family(j, ps, type{j}, scoring_fn, ns, discrete, data(:,u), params{j});
wolffd@0:   if verbose, fprintf('\nnode %d, empty score %6.4f\n', j, score); end
wolffd@0:   done = 0;
wolffd@0:   while ~done & (length(ps) <= max_fan_in)
wolffd@0:     pps = mysetdiff(order(1:i-1), ps); % potential parents
wolffd@0:     nps = length(pps);
wolffd@0:     pscore = zeros(1, nps);
wolffd@0:     for pi=1:nps
wolffd@0:       p = pps(pi);
wolffd@0:       pscore(pi) = score_family(j, [ps p], type{j}, scoring_fn, ns, discrete, data(:,u), params{j});
wolffd@0:       if verbose, fprintf('considering adding %d to %d, score %6.4f\n', p, j, pscore(pi)); end
wolffd@0:     end
wolffd@0:     [best_pscore, best_p] = max(pscore);
wolffd@0:     best_p = pps(best_p);
wolffd@0:     if best_pscore > score
wolffd@0:       score = best_pscore;
wolffd@0:       ps = [ps best_p];
wolffd@0:       if verbose, fprintf('* adding %d to %d, score %6.4f\n', best_p, j, best_pscore); end
wolffd@0:     else
wolffd@0:       done = 1;
wolffd@0:     end
wolffd@0:   end
wolffd@0:   if ~isempty(ps) % need this check for matlab 5.2
wolffd@0:     dag(ps, j) = 1;
wolffd@0:   end
wolffd@0: end
wolffd@0: 
wolffd@0: 
wolffd@0: 
wolffd@0: