loopy_dbn_inf_engine.m

来自「Bayes网络工具箱」· M 代码 · 共 80 行

function engine = loopy_dbn_inf_engine(bnet, onodes, max_iter, fast, momentum, obj)
% LOOPY_DBN_INF_ENGINE Loopy Pearl version of forwards-backwards
% engine = loopy_dbn_inf_engine(bnet, onodes, max_iter, fast, momentum, object_oriented)
%
% 'onodes' specifies which nodes are observed.
% 'max_iter' specifies the max num. forward-backward passes to perform (default: 1).
% 'fast' means pre-compute indices for table manipulation (as in jtree_fast)
% 'momentum' is as in loopy_pearl: default = 1 means use current msg only
% 'object_oriented' = 1 will call a method to compute msgs (good for noisy-or, but slow in general)
%
% The model must obey the same topological restrictions as hmm_inf_engine.
% In addition, each hidden node is assumed to have at most one observed child,
% and each observed child is assumed to have exactly one hidden parent.     
%
% For details of this algorithm, see
% "The Factored Frontier Algorithm for Approximate Inference in DBNs",
% Kevin Murphy and Yair Weiss, submitted to NIPS 2000.
%
% WARNING: THIS IS HIGHLY EXPERIMENTAL CODE!

if nargin < 3, max_iter = 1; end
if nargin < 4 | isempty(fast), fast = 0; else fast = 1; end
if nargin < 5, momentum = 0; end
if nargin < 6 | isempty(obj), obj = 0; else obj = 1; end

engine.object_oriented = obj;
engine.fast = fast;
engine.momentum = momentum;
engine.max_iter = max_iter;

engine.onodes = onodes;
engine.marginal = [];
engine.evidence = [];
engine.msg = [];
engine.parent_index = [];
engine.child_index = [];
%[engine.parent_index, engine.child_index] = mk_pearl_msg_indices(bnet); % need to unroll first

ss = length(bnet.intra);
hnodes = mysetdiff(1:ss, onodes);
obschild = zeros(1,ss);
for i=hnodes(:)'
  %ocs = myintersect(children(bnet.dag, i), onodes);
  ocs = children(bnet.intra, i);
  assert(length(ocs) <= 1);
  if length(ocs)==1
    obschild(i) = ocs(1);
  end
end
engine.obschild = obschild;

engine.mult_self_ndx = [];
engine.mult_parent_ndx = [];
engine.marg_self_ndx = [];
engine.marg_parent_ndx = [];


if fast
  onodes2 = [onodes(:); onodes(:)+ss];
  ns = bnet.node_sizes(:);
  ns(onodes2) = 1;
  engine.mult_self_ndx = cell(1, ss);
  engine.mult_parent_ndx = cell(ss, ss);
  engine.marg_self_ndx = cell(1, ss);
  engine.marg_parent_ndx = cell(ss, ss);
  for i=hnodes(:)'
    ps = parents(bnet.inter, i);
    fam = [ps i+ss];
    engine.mult_self_ndx{i} = mk_multiply_table_ndx(fam, i+ss, ns);
    engine.marg_self_ndx{i} = mk_marginalise_table_ndx(fam, i+ss, ns);
    for p=ps(:)'
      engine.mult_parent_ndx{i,p} = mk_multiply_table_ndx(fam, p, ns);
      engine.marg_parent_ndx{i,p} = mk_marginalise_table_ndx(fam, p, ns);
    end
  end
end

engine = class(engine, 'loopy_dbn_inf_engine', inf_engine(bnet));