MCMC方法的关键是通过构造平稳分布为P的马尔科夫链来产生样本。在贝叶斯网络中,产生的样本就是各个贝叶斯结构,通过产生的样本(这些结构)中,选取可行的结构。关键的部分就在于如何仅仅通过训练集(每个节点一系列状态),来得到采样样本(最终结构)。
MH算法是MCMC算法的重要代表,MH是通过上一轮采样结果来采样,获得候选样本,但是这个样本有可能被拒绝掉,为什么会被拒绝掉呢?因为算法中规定了一个U是在0-1范围内的随机数,如果接受概率大于它则接受。最终,,,(
这个拒绝的步骤,在我理解看来,是为了达到平稳分布的一个构造过程。由于在x*最终达到平稳状态时:细致平稳的要求是
(1)
即达到这样的效果:
P转移到q和q转移到p的量是一样的,那么这个马尔科夫链是平稳的。
那么,由于一般的马尔科夫链不是平稳的,为了构造一个平稳的马尔科夫链,需要引入和一个接受概率。
如下所说:构造了一个接受率,那么马尔科夫链达到了平稳状态。
但是接受率的写法为甚是
其实这也很简单:
这个算法被应用在了贝叶斯结构学习中,通过贝叶斯工具箱,其中有一个learn_struct_mcmc.m函数,输入数据为训练集,每个节点可取的状态值,输出为采样的结构,接受率以及第t次迭代的边数。代码如下:
function [sampled_graphs, accept_ratio, num_edges] = learn_struct_mcmc(data, ns, varargin) % MY_LEARN_STRUCT_MCMC Monte Carlo Markov Chain search over DAGs assuming fully observed data % [sampled_graphs, accept_ratio, num_edges] = learn_struct_mcmc(data, ns, ...) % % data(i,m) is the value of node i in case m. % ns(i) is the number of discrete values node i can take on. % % sampled_graphs{m} is the m'th sampled graph. % accept_ratio(t) = acceptance ratio at iteration t % num_edges(t) = number of edges in model at iteration t % % The following optional arguments can be specified in the form of name/value pairs: % [default value in brackets] % % scoring_fn - 'bayesian' or 'bic' [ 'bayesian' ] % Currently, only networks with all tabular nodes support Bayesian scoring. % type - type{i} is the type of CPD to use for node i, where the type is a string % of the form 'tabular', 'noisy_or', 'gaussian', etc. [ all cells contain 'tabular' ] % params - params{i} contains optional arguments passed to the CPD constructor for node i, % or [] if none. [ all cells contain {'prior', 1}, meaning use uniform Dirichlet priors ] % discrete - the list of discrete nodes [ 1:N ] % clamped - clamped(i,m) = 1 if node i is clamped in case m [ zeros(N, ncases) ] % nsamples - number of samples to draw from the chain after burn-in [ 100*N ] % burnin - number of steps to take before drawing samples [ 5*N ] % init_dag - starting point for the search [ zeros(N,N) ] % % e.g., samples = my_learn_struct_mcmc(data, ns, 'nsamples', 1000); % % Modified by Sonia Leach (SML) 2/4/02, 9/5/03
[n ncases] = size(data);%n是节点数,ncase是样本数
% set default params type = cell(1,n); params = cell(1,n);%定义类型和参数 for i=1:n type{i} = 'tabular'; %params{i} = { 'prior', 1}; params{i} = { 'prior_type', 'dirichlet', 'dirichlet_weight', 1 }; end scoring_fn = 'bayesian'; discrete = 1:n; clamped = zeros(n, ncases);%定义一个和训练集一样大的零矩阵 nsamples = 100*n; burnin = 5*n; dag = zeros(n);
args = varargin;%arg是可变参数列表 nargs = length(args); for i=1:2:nargs switch args{i}, case'nsamples', nsamples = args{i+1}; case'burnin', burnin = args{i+1}; case'init_dag', dag = args{i+1}; case'scoring_fn', scoring_fn = args{i+1}; case'type', type = args{i+1}; case'discrete', discrete = args{i+1}; case'clamped', clamped = args{i+1}; case'params', if isempty(args{i+1}), params = cell(1,n); else params = args{i+1}; end end end
% We implement the fast acyclicity check described by P. Giudici and R. Castelo, % "Improving MCMC model search for data mining", submitted to J. Machine Learning, 2001.
% SML: also keep descendant matrix C use_giudici = 1; if use_giudici [nbrs, ops, nodes, A] = mk_nbrs_of_digraph(dag); else [nbrs, ops, nodes] = mk_nbrs_of_dag(dag); A = []; end
num_accepts = 1; num_rejects = 1; T = burnin + nsamples; accept_ratio = zeros(1, T);%定义接受率矩阵 num_edges = zeros(1, T);%边数目的矩阵 sampled_graphs = cell(1, nsamples);%采样图 %sampled_bitv = zeros(nsamples, n^2);
for t=1:T %对总共的点数,进行take_step操作,得到accept [dag, nbrs, ops, nodes, A, accept] = take_step(dag, nbrs, ops, ... nodes, ns, data, clamped, A, ... scoring_fn, discrete, type, params); num_edges(t) = sum(dag(:)); num_accepts = num_accepts + accept;%接受数累加 num_rejects = num_rejects + (1-accept);%拒绝数累加 accept_ratio(t) = num_accepts/num_rejects;%重复更新接受率 if t > burnin%如t超出了舍弃范围 sampled_graphs{t-burnin} = dag;%把图放进样本图中去。 %sampled_bitv(t-burnin, :) = dag(:)'; end end
%%%%%%%%%
function [new_dag, new_nbrs, new_ops, new_nodes, A, accept] = ... take_step(dag, nbrs, ops, nodes, ns, data, clamped, A, ... scoring_fn, discrete, type, params, prior_w)
use_giudici = ~isempty(A); if use_giudici %如果矩阵A是非空,更新A [new_dag, op, i, j, new_A] = pick_digraph_nbr(dag, nbrs, ops, nodes,A); % updates A [new_nbrs, new_ops, new_nodes] = mk_nbrs_of_digraph(new_dag, new_A); else d = sample_discrete(normalise(ones(1, length(nbrs)))); new_dag = nbrs{d}; op = ops{d}; i = nodes(d, 1); j = nodes(d, 2); [new_nbrs, new_ops, new_nodes] = mk_nbrs_of_dag(new_dag); end
bf = bayes_factor(dag, new_dag, op, i, j, ns, data, clamped, scoring_fn, discrete, type, params);%bf是一个什么值?
%R = bf * (new_prior / prior) * (length(nbrs) / length(new_nbrs)); R = bf * (length(nbrs) / length(new_nbrs)); u = rand(1,1); if u > min(1,R) % reject the move 拒绝采样 accept = 0; new_dag = dag; new_nbrs = nbrs; new_ops = ops; new_nodes = nodes; else accept = 1;%接受采样的话,对A进行更新 if use_giudici A = new_A; % new_A already updated in pick_digraph_nbr end end
%%%%%%%%%
function bfactor = bayes_factor(old_dag, new_dag, op, i, j, ns, data, clamped, scoring_fn, discrete, type, params)
u = find(clamped(j,:)==0); LLnew = score_family(j, parents(new_dag, j), type{j}, scoring_fn, ns, discrete, data(:,u), params{j}); LLold = score_family(j, parents(old_dag, j), type{j}, scoring_fn, ns, discrete, data(:,u), params{j}); bf1 = exp(LLnew - LLold);%新得分-旧得分取指数
if strcmp(op, 'rev') % must also multiply in the changes to i's family u = find(clamped(i,:)==0); LLnew = score_family(i, parents(new_dag, i), type{i}, scoring_fn, ns, discrete, data(:,u), params{i}); LLold = score_family(i, parents(old_dag, i), type{i}, scoring_fn, ns, discrete, data(:,u), params{i}); bf2 = exp(LLnew - LLold); else bf2 = 1; end bfactor = bf1 * bf2;
%%%%%%%% Giudici stuff follows %%%%%%%%%%
% SML: This now updates A as it goes from digraph it choses function [new_dag, op, i, j, new_A] = pick_digraph_nbr(dag, digraph_nbrs, ops, nodes, A)
d = sample_discrete(normalise(ones(1, length(digraph_nbrs)))); %d = myunidrnd(length(digraph_nbrs),1,1); i = nodes(d, 1); j = nodes(d, 2); new_dag = digraph_nbrs(:,:,d); op = ops{d}; new_A = update_ancestor_matrix(A, op, i, j, new_dag);
%%%%%%%%%%%%%%
% 这是对结构的三种操作: function A = update_ancestor_matrix(A, op, i, j, dag)
switch op case'add', A = do_addition(A, op, i, j, dag); case'del', A = do_removal(A, op, i, j, dag); case'rev', A = do_removal(A, op, i, j, dag); A = do_addition(A, op, j, i, dag); end
%%%%%%%%%%%% % 这是加边操作: function A = do_addition(A, op, i, j, dag)
A(j,i) = 1; % i is an ancestor of j anci = find(A(i,:)); if ~isempty(anci) A(j,anci) = 1; % all of i's ancestors are added to Anc(j) end ancj = find(A(j,:)); descj = find(A(:,j)); if ~isempty(ancj) for k=descj(:)' A(k,ancj) = 1; % all of j's ancestors are added to each descendant of j end end
%%%%%%%%%%%这是剪边操作 function A = do_removal(A, op, i, j, dag)
% find all the descendants of j, and put them in topological order
% SML: originally Kevin had the next line commented and the %* lines % being used but I think this is equivalent and much less expensive % I assume he put it there for debugging and never changed it back...? descj = find(A(:,j)); %* R = reachability_graph(dag); %* descj = find(R(j,:));
order = topological_sort(dag);
% SML: originally Kevin used the %* line but this was extracting the % wrong things to sort %* descj_topnum = order(descj); [junk, perm] = sort(order); %SML:node i is perm(i)-TH in order descj_topnum = perm(descj); %SML:descj(i) is descj_topnum(i)-th in order
% SML: now re-sort descj by rank in descj_topnum [junk, perm] = sort(descj_topnum); descj = descj(perm);
% Update j and all its descendants A = update_row(A, j, dag); for k = descj(:)' A = update_row(A, k, dag); end
%%%%%%%%%%%
function A = old_do_removal(A, op, i, j, dag)
% find all the descendants of j, and put them in topological order % SML: originally Kevin had the next line commented and the %* lines % being used but I think this is equivalent and much less expensive % I assume he put it there for debugging and never changed it back...? descj = find(A(:,j)); %* R = reachability_graph(dag); %* descj = find(R(j,:));
order = topological_sort(dag); descj_topnum = order(descj); [junk, perm] = sort(descj_topnum); descj = descj(perm); % Update j and all its descendants A = update_row(A, j, dag); for k = descj(:)' A = update_row(A, k, dag); end
%%%%%%%%%升级
function A = update_row(A, j, dag)
% We compute row j of A A(j, :) = 0; ps = parents(dag, j); if ~isempty(ps) A(j, ps) = 1; end for k=ps(:)' anck = find(A(k,:)); if ~isempty(anck) A(j, anck) = 1; end end
%%%%%%%%
function A = init_ancestor_matrix(dag)
order = topological_sort(dag); A = zeros(length(dag)); for j=order(:)' A = update_row(A, j, dag); end
|
总结一下,再说说我的疑惑,MH算法是用列举完全的方法把所有的边操作列举出来,然后再对这个集合进行采样吗?当节点数目多的时候怎么办呢?这是O(n^2)数量级的数目吧。那么是否有好的办法可以改进此抽样算法?
若用MCMC采样,需要对每一条边进行采样,当有N个节点,则有N*(N-1)/2条边需要被采样,每条边是3个值,分别代表顺、逆、无边。而组合出来的结构种类数目一共是(3^(n*(n-1)/2)),当然这里没有考虑去除环的结构。
在《基于MCMC贝叶斯网络学习算法》中提到,在构造马尔科夫链中,一个结构G和一个对G改变了一条边的集合nbr(G),是马尔科夫链的两个状态。此文中相对于经典MH算法的改进在于:由于搜索空间巨大,先利用条件互信息,条件独立性测试进行了节点相互位置的固定,再对边进行添加,反转,(没有无边状态,由于独立性测试已经确定依赖关系)。最后此方法学习结果和原来的基本一致,接受率也一致,但是文中也并没有提到新算法是否节约了运算时间?优点没有实际的例子做支撑,私下觉得略有一点草率。另,这篇文章的目标是把此算法用在文档分类中,这个可以再深入研究一下,应该属于自然语言处理范围,也是概率图模型的一种主要应用场景了。