smooth_evidence.m

贝叶斯网络的matlab实现。可以创建贝叶斯网络、训练模型

function [marginal, msg, loglik] = smooth_evidence(engine, evidence)

[ss T] = size(evidence);
bnet = bnet_from_engine(engine);
onodes = engine.onodes;
hnodes = mysetdiff(1:ss, onodes);
hnodes = hnodes(:)';

ns = bnet.node_sizes(:);
onodes2 = [onodes(:); onodes(:)+ss];
ns(onodes2) = 1;
	   
verbose = 0;
pot_type = 'd';
niter = engine.max_iter;

if verbose, fprintf('new smooth\n'); end

% msg(i1,t1,i2,j2) (i1,t1) -> (i2,t2)
%lambda_msg = cell(ss,T,ss,T);
%pi_msg = cell(ss,T,ss,T);

% intra_lambda_msg(i,j,t) (i,t) -> (j,t), i is child
% inter_lambda_msg(i,j,t) (i,t+1) -> (j,t), i is child
% inter_pi_msg(i,j,t) (i,t-1) -> (j,t), i is parent
intra_lambda_msg = cell(ss,ss,T);
inter_lambda_msg = cell(ss,ss,T);
inter_pi_msg = cell(ss,ss,T);

lambda = cell(ss,T);
pi = cell(ss,T);

for t=1:T
  for i=1:ss
    lambda{i,t} = ones(ns(i), 1);
    pi{i,t} = ones(ns(i), 1);
    
    cs = children(bnet.intra, i);
    for c=cs(:)'
      intra_lambda_msg{c,i,t} = ones(ns(i),1);
    end
    
    cs = children(bnet.inter, i);
    for c=cs(:)'
      inter_lambda_msg{c,i,t} = ones(ns(i),1);
    end
    
    ps = parents(bnet.inter, i);
    for p=ps(:)'
      inter_pi_msg{p,i,t} = ones(ns(i), 1); % not used for t==1
    end
  end
end


% each hidden node absorbs lambda from its observed child (if any)
for t=1:T
  for i=hnodes
    c = engine.obschild(i);
    if c > 0
      if t==1
	fam = family(bnet.dag, c);
	e = bnet.equiv_class(c, 1);
	CPDpot = CPD_to_pot(pot_type, bnet.CPD{e}, fam, bnet.node_sizes(:), bnet.cnodes(:), evidence(:,1));
      else
	fam = family(bnet.dag, c, 2); % within 2 slice network
	e = bnet.equiv_class(c, 2);
	CPDpot = CPD_to_pot(pot_type, bnet.CPD{e}, fam, bnet.node_sizes(:), bnet.cnodes(:), evidence(:,t-1:t));
      end
      temp = pot_to_marginal(CPDpot);
      lam_msg = normalise(temp.T);
      intra_lambda_msg{c,i,t} = lam_msg;
    end
  end
end

for iter=1:engine.max_iter
  % FORWARD
  for t=1:T
    % update pi
    for i=hnodes
      if t==1
	e = bnet.equiv_class(i,1);
	CPD = struct(bnet.CPD{e});
	pi{i,t} = CPD.CPT;
      else
	e = bnet.equiv_class(i,2);
	CPD = struct(bnet.CPD{e});
	ps = parents(bnet.inter, i);
	dom = [ps i+ss];
	pot = dpot(dom, ns(dom), CPD.CPT);
	for p=ps(:)'
	  temp = dpot(p, ns(p), inter_pi_msg{p,i,t});
	  pot = multiply_by_pot(pot, temp);
	end
	pot = marginalize_pot(pot, i+ss);
	temp = pot_to_marginal(pot);
	pi{i,t} = temp.T;
      end
      if verbose, fprintf('%d updates pi\n', i+(t-1)*ss); disp(pi{i,t}); end
    end
    
    % send pi msg to children 
    for i=hnodes
      cs = children(bnet.inter, i);
      for c=cs(:)'
	pot = pi{i,t};
	for k=cs(:)'
	  if k ~= c
	    pot = pot .* inter_lambda_msg{k,i,t};
	  end
	end
	cs2 = children(bnet.intra, i);
	for k=cs2(:)'
	  pot = pot .* intra_lambda_msg{k,i,t};
	end
	inter_pi_msg{i,c,t+1} = normalise(pot);
	if verbose, fprintf('%d sends pi to %d\n', i+(t-1)*ss, c+t*ss); disp(inter_pi_msg{i,c,t+1}); end
      end
    end
  end

  if verbose, fprintf('backwards\n'); end
  % BACKWARD
  for t=T:-1:1
    % update lambda
    for i=hnodes
      pot = ones(ns(i), 1);
      cs = children(bnet.inter, i);
      for c=cs(:)'
	pot = pot .* inter_lambda_msg{c,i,t};
      end
      cs = children(bnet.intra, i);
      for c=cs(:)'
	pot = pot .* intra_lambda_msg{c,i,t};
      end
      lambda{i,t} = normalise(pot);
      if verbose, fprintf('%d computes lambda\n', i+(t-1)*ss); disp(lambda{i,t}); end
    end
    
    % send lambda msgs to hidden parents in prev slcie
    for i=hnodes
      ps = parents(bnet.inter, i);
      if t > 1
	e = bnet.equiv_class(i, 2);
	CPD = struct(bnet.CPD{e});
	fam = [ps i+ss];
	for p=ps(:)'
	  pot = dpot(fam, ns(fam), CPD.CPT);
	  temp = dpot(i+ss, ns(i), lambda{i,t});
	  pot = multiply_by_pot(pot, temp);
	  for k=ps(:)'
	    if k ~= p
	      temp = dpot(k, ns(k), inter_pi_msg{k,i,t});
	      pot = multiply_by_pot(pot, temp);
	    end
	  end
	  pot = marginalize_pot(pot, p);
	  temp = pot_to_marginal(pot);
	  inter_lambda_msg{i,p,t-1} = normalise(temp.T);
	  if verbose, fprintf('%d sends lambda to %d\n', i+(t-1)*ss, p+(t-2)*ss); disp(inter_lambda_msg{i,p,t-1}); end
	end
      end
    end
  end
end



marginal = cell(ss,T);
for t=1:T
  for i=hnodes
    marginal{i,t} = normalise(pi{i,t} .* lambda{i,t});     
  end
end

loglik = 0;

msg.inter_pi_msg = inter_pi_msg;
msg.inter_lambda_msg = inter_lambda_msg;
msg.intra_lambda_msg = intra_lambda_msg;