camir-aes2014: toolboxes/FullBNT-1.0.7/bnt/learning/learn_params

first hg version after svn

rev	line source
wolffd@0	1 function bnet = learn_params_dbn(bnet, data)
wolffd@0	2 % LEARN_PARAM_DBN Estimate params of a DBN for a fully observed model
wolffd@0	3 % bnet = learn_params_dbn(bnet, data)
wolffd@0	4 %
wolffd@0	5 % data(i,t) is the value of node i in slice t (can be a cell array)
wolffd@0	6 % We currently assume there is a single time series
wolffd@0	7 %
wolffd@0	8 % We set bnet.CPD{i} to its ML/MAP estimate.
wolffd@0	9 %
wolffd@0	10 % Currently we assume each node in the first 2 slices has its own CPD (no param tying);
wolffd@0	11 % all nodes in slices >2 share their params with slice 2 as usual.
wolffd@0	12
wolffd@0	13 [ss T] = size(data);
wolffd@0	14
wolffd@0	15 % slice 1
wolffd@0	16 for j=1:ss
wolffd@0	17 if adjustable_CPD(bnet.CPD{j})
wolffd@0	18 fam = family(bnet.dag,j);
wolffd@0	19 bnet.CPD{j} = learn_params(bnet.CPD{j}, data(fam,1));
wolffd@0	20 end
wolffd@0	21 end
wolffd@0	22
wolffd@0	23
wolffd@0	24 % slices 2:T
wolffd@0	25 % data2(:,t) contains [data(:,t-1); data(:,t)].
wolffd@0	26 % Then we extract out the rows corresponding to the parents in the current and previous slice.
wolffd@0	27 data2 = [data(:,1:T-1);
wolffd@0	28 data(:,2:T)];
wolffd@0	29 for j=1:ss
wolffd@0	30 j2 = j+ss;
wolffd@0	31 if adjustable_CPD(bnet.CPD{j2})
wolffd@0	32 fam = family(bnet.dag,j2);
wolffd@0	33 bnet.CPD{j2} = learn_params(bnet.CPD{j2}, data2(fam,:));
wolffd@0	34 end
wolffd@0	35 end
wolffd@0	36

Mercurial > hg > camir-aes2014