loopy_pearl_inf_engine.m

来自「Bayes网络工具箱」· M 代码 · 共 46 行

function engine = loopy_pearl_inf_engine(bnet, max_iter, tol, momentum)
% LOOPY_PEARL_INF_ENGINE Pearl's algorithm (belief propagation) for graphs which may contain loops
% engine = loopy_pearl_inf_engine(bnet, max_iter, tol, momentum)
%
% We run loopy propagation for a maximum of 'max_iter' iterations, or until
% all the posteriors differ by less than 'tol', whichever occurs first.
% 'max_iter' defaults to the num. nodes in the graph, which is an upper bound on the radius.
% If the graph has no loops (undirected cycles), the results will be exact, because by then every
% node is guaranteed to have felt the influence of all other nodes.
% 'tol' defaults to 1e-3.
% If 'momentum' is specified, the messages that are sent will be a convex combination of
% the current and previous messages (m*old + (1-m)*new). Default: momentum = 0.
% 
% For some theoretical work on the properties of this algorithm, see
% - "Correctness of local probability propagation in graphical models with loops",
%       Yair Weiss, Neural Computation, 12 (1-41) 2000.
% - "Correctness of belief propagation in Gaussian graphical models of arbitrary topology",
%     Y. Weiss and W. T. Freeman, TR UCB--CSD-99-1046 
% For some experimental work, see
% - "Loopy belief propagation for approximate inference: an empirical study",
%      K. Murphy, Y. Weiss and M. Jordan, UAI 99.
%
% SEE ALSO pearl_inf_engine


N = length(bnet.dag);

if nargin < 2, max_iter = N; end
if nargin < 3, tol = 1e-3; end
if nargin < 4, momentum = 0; end

engine.max_iter = max_iter;
engine.tol = tol;
engine.momentum = momentum;

% this is where we store stuff between enter_evidence and marginal_nodes
engine.marginal = cell(1,N);
engine.evidence = []; 
engine.msg = [];

[engine.parent_index, engine.child_index] = mk_pearl_msg_indices(bnet);

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