jtree_dbn_inf_engine.m

来自「Bayes网络工具箱」· M 代码 · 共 26 行

function engine = jtree_dbn_inf_engine(bnet, onodes)
% JTREE_DBN_INF_ENGINE Junction tree algo for variable length DBNs
% engine = jtree_dbn_inf_engine(bnet, onodes)
%
% 'onodes' is an optional specification which nodes are observed nodes in slice 1, 
% which is assumed to be repeated across all slices. It may be used to optimize the jtree.
%
% This engine is just "syntactic sugar" for the exact version of BK, i.e., it is equivalent to
%     engine = bk_inf_engine(bnet, 'exact', onodes)
% This may change in the future.
%
% Note that BK only ever applies jtree to the 2-slice network.  An alternative would be to apply jtree
% to the unrolled network.  This is what 'jtree_unrolled_dbn_inf_engine' does.  Unfortunately,
% computing the jtree can be slow for long sequences, and this does not support online inference.
% Instead we could try to find the repeating pattern in the jtree and unroll the jtree itself whenever
% the sequence length changes. (Note: the jtree will only have a repeating pattern if we constrain the
% elimination order to respect the temporal order.)  This can be fast, but uses a lot of space,
% because we keep all the intermediate cliques which are used to advance the belief state by one time step.


%parent
pengine = bk_inf_engine(bnet, 'exact', onodes);

engine.onodes = onodes;
engine = class(engine, 'jtree_dbn_inf_engine', pengine);