Daniel@0: % We compare MCMC structure learning with exhaustive enumeration of all dags.
Daniel@0: 
Daniel@0: N = 3;
Daniel@0: %N = 4;
Daniel@0: dag = mk_rnd_dag(N);
Daniel@0: ns = 2*ones(1,N);
Daniel@0: bnet = mk_bnet(dag, ns);
Daniel@0: for i=1:N
Daniel@0:   bnet.CPD{i} = tabular_CPD(bnet, i);
Daniel@0: end
Daniel@0: 
Daniel@0: ncases = 100;
Daniel@0: data = zeros(N, ncases);
Daniel@0: for m=1:ncases
Daniel@0:   data(:,m) = cell2num(sample_bnet(bnet));
Daniel@0: end
Daniel@0: 
Daniel@0: dags = mk_all_dags(N);
Daniel@0: score = score_dags(data, ns, dags);
Daniel@0: post  = normalise(exp(score));
Daniel@0: 
Daniel@0: [sampled_graphs, accept_ratio] = learn_struct_mcmc(data, ns, 'nsamples', 100, 'burnin', 10);
Daniel@0: mcmc_post = mcmc_sample_to_hist(sampled_graphs, dags);
Daniel@0: 
Daniel@0: if 0
Daniel@0:   subplot(2,1,1)
Daniel@0:   bar(post)
Daniel@0:   subplot(2,1,2)
Daniel@0:   bar(mcmc_post)
Daniel@0:   print(gcf, '-djpeg', '/home/cs/murphyk/public_html/Bayes/Figures/mcmc_post.jpg')
Daniel@0: 
Daniel@0:   clf
Daniel@0:   plot(accept_ratio)
Daniel@0:   print(gcf, '-djpeg', '/home/cs/murphyk/public_html/Bayes/Figures/mcmc_accept.jpg')
Daniel@0: end