008 gtsam/examples/HMMExample.cpp

隐马尔可夫模型(离散Bayes Net)

一、main函数

构建Values和Ordering

  // Define variables as well as ordering
  Ordering ordering;
  vector<DiscreteKey> keys;
  for (int k = 0; k < nrNodes; k++) {
    DiscreteKey key_i(k, nrStates);
    keys.push_back(key_i);
    ordering.emplace_back(k);
  }

创建离散的Bayes Net作为模型

  // Create HMM as a DiscreteBayesNet
  DiscreteBayesNet hmm;

重载了 | 操作符

Helper function to create Signature objects
example: Signature s = D | E;

定义HMM和加入观测

  // Define backbone
  const string transition = "8/1/1 1/8/1 1/1/8";
  for (int k = 1; k < nrNodes; k++) {
    hmm.add(keys[k] | keys[k - 1] = transition);
  }

  // Add some measurements, not needed for all time steps!
  hmm.add(keys[0] % "7/2/1");
  hmm.add(keys[1] % "1/9/0");
  hmm.add(keys.back() % "5/4/1");
  // print
  hmm.print("HMM");

转换为因子图并求解消元
按时间箭头反序排列

  // Convert to factor graph
  DiscreteFactorGraph factorGraph(hmm);

  // Create solver and eliminate
  // This will create a DAG ordered with arrow of time reversed
  DiscreteBayesNet::shared_ptr chordal =
      factorGraph.eliminateSequential(ordering);
  chordal->print("Eliminated");

  // solve
  DiscreteFactor::sharedValues mpe = chordal->optimize();
  GTSAM_PRINT(*mpe);

sampling

  // We can also sample from it
  cout << "\n10 samples:" << endl;
  for (size_t k = 0; k < 10; k++) {
    DiscreteFactor::sharedValues sample = chordal->sample();
    GTSAM_PRINT(*sample);
  }

获取边际概率

  // Or compute the marginals. This re-eliminates the FG into a Bayes tree
  cout << "\nComputing Node Marginals .." << endl;
  DiscreteMarginals marginals(factorGraph);
  for (int k = 0; k < nrNodes; k++) {
    Vector margProbs = marginals.marginalProbabilities(keys[k]);
    stringstream ss;
    ss << "marginal " << k;
    print(margProbs, ss.str());
  }