A Quick Tour of the Boost Graph Library

The domain of graph data structures and algorithms is in some respects more complicated than that of containers. The abstract iterator interface used by STL is not sufficiently rich to encompass the numerous ways that graph algorithms may traverse a graph. Instead, we formulate an abstract interface that serves the same purpose for graphs that iterators do for basic containers (though iterators still play a large role). Figure 1 depicts the analogy between the STL and the BGL.

Figure 1: The analogy between the STL and the BGL.

Figure 2: An example of a directed graph.

In the following sections we will use the BGL to construct this example graph and manipulate it in various ways. The complete source code for this example can be found in examples/quick_tour.cpp. Each of the following sections discusses a "slice" of this example file. Excerpts from the output of the example program will also be listed.

Constructing a Graph

In this example we will use the BGL adjacency_list class to demonstrate the main ideas in the BGL interface. The adjacency_list class provides a generalized version of the classic "adjacency list" data structure. The adjacency_list is a template class with six template parameters, though here we only fill in the first three parameters and use the defaults for the remaining three. The first two template arguments (vecS, vecS) determine the data structure used to represent the out-edges for each vertex in the graph and the data structure used to represent the graph's vertex set (see section Choosing the Edgelist and VertexList for information about the tradeoffs of the different data structures). The third argument, bidirectionalS, selects a directed graph that provides access to both out and in-edges. The other options for the third argument are directedS which selects a directed graph with only out-edges, and undirectedS which selects an undirected graph.

Once we have the graph type selected, we can create the graph in Figure 2 by declaring a graph object and filling in edges using the add_edge() function of the MutableGraph interface (which adjacency_list implements). We use the array of pairs edge_array merely as a convenient way to explicitly create the edges for this example.

  #include <iostream>                  // for std::cout
  #include <utility>                   // for std::pair
  #include <algorithm>                 // for std::for_each
  #include <boost/graph/graph_traits.hpp>
  #include <boost/graph/adjacency_list.hpp>
  #include <boost/graph/dijkstra_shortest_paths.hpp>

  using namespace boost;

  int main(int,char*[])
  {
    // create a typedef for the Graph type
    typedef adjacency_list<vecS, vecS, bidirectionalS> Graph;

    // Make convenient labels for the vertices
    enum { A, B, C, D, E, N };
    const int num_vertices = N;
    const char* name = "ABCDE";

    // writing out the edges in the graph
    typedef std::pair<int, int> Edge;
    Edge edge_array[] =
    { Edge(A,B), Edge(A,D), Edge(C,A), Edge(D,C),
      Edge(C,E), Edge(B,D), Edge(D,E) };
    const int num_edges = sizeof(edge_array)/sizeof(edge_array[0]);

    // declare a graph object
    Graph g(num_vertices);

    // add the edges to the graph object
    for (int i = 0; i < num_edges; ++i)
      add_edge(edge_array[i].first, edge_array[i].second, g);
    ...
    return 0;
  }

Instead of calling the add_edge() function for each edge, we could use the edge iterator constructor of the graph. This is typically more efficient than using add_edge(). Pointers to the edge_array can be viewed as iterators, so we can call the iterator constructor by passing pointers to the beginning and end of the array.

    Graph g(edge_array, edge_array + sizeof(edge_array) / sizeof(Edge), num_vertices);

Instead of creating a graph with a certain number of vertices to begin with, it is also possible to add and remove vertices with the add_vertex() and remove_vertex() functions, also of the MutableGraph interface.

Accessing the Vertex Set

Now that we have created a graph, we can use the graph interface to access the graph data in different ways. First we can access all of the vertices in the graph using the vertices() function of the VertexListGraph interface. This function returns a std::pair of vertex iterators (the first iterator points to the "beginning" of the vertices and the second iterator points "past the end"). Dereferencing a vertex iterator gives a vertex object. The type of the vertex iterator is given by the graph_traits class. Note that different graph classes can have different associated vertex iterator types, which is why we need the graph_traits class. Given some graph type, the graph_traits class will provide access to the vertex_iterator type.

The following example prints out the index for each of the vertices in the graph. All vertex and edge properties, including index, are accessed via property map objects. The property_map class is used to obtain the property map type for a specific property (specified by vertex_index_t, one of the BGL predefined properties) and function call get(vertex_index, g) returns the actual property map object.

  // ...
  int main(int,char*[])
  {
    // ...

    typedef graph_traits<Graph>::vertex_descriptor Vertex;

    // get the property map for vertex indices
    typedef property_map<Graph, vertex_index_t>::type IndexMap;
    IndexMap index = get(vertex_index, g);

    std::cout << "vertices(g) = ";
    typedef graph_traits<Graph>::vertex_iterator vertex_iter;
    std::pair<vertex_iter, vertex_iter> vp;
    for (vp = vertices(g); vp.first != vp.second; ++vp.first) {
      Vertex v = *vp.first;
      std::cout << index[v] <<  " ";
    }
    std::cout << std::endl;
    // ...
    return 0;
  }

  vertices(g) = 0 1 2 3 4

Accessing the Edge Set

The set of edges for a graph can be accessed with the edges() function of the EdgeListGraph interface. Similar to the vertices() function, this returns a pair of iterators, but in this case the iterators are edge iterators. Dereferencing an edge iterator gives an edge object. The source() and target() functions return the two vertices that are connected by the edge. Instead of explicitly creating a std::pair for the iterators, this time we will use the boost::tie() helper function. This handy function can be used to assign the parts of a std::pair into two separate variables, in this case ei and ei_end. This is usually more convenient than creating a std::pair and is our method of choice for the BGL.

  // ...
  int main(int,char*[])
  {
    // ...
    std::cout << "edges(g) = ";
    graph_traits<Graph>::edge_iterator ei, ei_end;
    for (boost::tie(ei, ei_end) = edges(g); ei != ei_end; ++ei)
        std::cout << "(" << index[source(*ei, g)]
                  << "," << index[target(*ei, g)] << ") ";
    std::cout << std::endl;
    // ...
    return 0;
  }

  edges(g) = (0,1) (0,3) (2,0) (3,2) (2,4) (1,3) (3,4)

The Adjacency Structure

In the next few examples we will explore the adjacency structure of the graph from the point of view of a particular vertex. We will look at the vertices' in-edges, out-edges, and its adjacent vertices. We will encapsulate this in an "exercise vertex" function, and apply it to each vertex in the graph. To demonstrate the STL-interoperability of BGL, we will use the STL for_each() function to iterate through the vertices and apply the function.

  //...
  int main(int,char*[])
  {
    //...
    std::for_each(vertices(g).first, vertices(g).second,
                  exercise_vertex<Graph>(g));
    return 0;
  }

We use a functor for exercise_vertex instead of just a function because the graph object will be needed when we access information about each vertex; using a functor gives us a place to keep a reference to the graph object during the execution of the std::for_each(). Also we template the functor on the graph type so that it is reusable with different graph classes. Here is the start of the exercise_vertex functor:

  template <class Graph> struct exercise_vertex {
    exercise_vertex(Graph& g_) : g(g_) {}
    //...
    Graph& g;
  };

Vertex Descriptors

The first thing we need to know in order to write the operator() method of the functor is the type for the vertex objects of the graph. The vertex type will be the parameter to the operator() method. To be precise, we do not deal with actual vertex objects, but rather with vertex descriptors. Many graph representations (such as adjacency lists) do not store actual vertex objects, while others do (e.g., pointer-linked graphs). This difference is hidden underneath the "black-box" of the vertex descriptor object. The vertex descriptor is something provided by each graph type that can be used to access information about the graph via the out_edges(), in_edges(), adjacent_vertices(), and property map functions that are described in the following sections. The vertex_descriptor type is obtained through the graph_traits class. The typename keyword used below is necessary because the type on the left hand side of the scope :: operator (the graph_traits<Graph> type) is dependent on a template parameter (the Graph type). Here is how we define the functor's apply method:

  template <class Graph> struct exercise_vertex {
    //...
    typedef typename graph_traits<Graph>
      ::vertex_descriptor Vertex;

    void operator()(const Vertex& v) const
    {
      //...
    }
    //...
  };

Out-Edges, In-Edges, and Edge Descriptors

The out-edges of a vertex are accessed with the out_edges() function of the IncidenceGraph interface. The out_edges() function takes two arguments: the first argument is the vertex and the second is the graph object. The function returns a pair of iterators which provide access to all of the out-edges of a vertex (similar to how the vertices() function returned a pair of iterators). The iterators are called out-edge iterators and dereferencing one of these iterators gives an edge descriptor object. An edge descriptor plays the same kind of role as the vertex descriptor object, it is a "black box" provided by the graph type. The following code snippet prints the source-target pairs for each out-edge of vertex v.

  template <class Graph> struct exercise_vertex {
    //...
    void operator()(const Vertex& v) const
    {
      typedef graph_traits<Graph> GraphTraits;
      typename property_map<Graph, vertex_index_t>::type
        index = get(vertex_index, g);

      std::cout << "out-edges: ";
      typename GraphTraits::out_edge_iterator out_i, out_end;
      typename GraphTraits::edge_descriptor e;
      for (boost::tie(out_i, out_end) = out_edges(v, g);
           out_i != out_end; ++out_i) {
        e = *out_i;
        Vertex src = source(e, g), targ = target(e, g);
        std::cout << "(" << index[src] << ","
                  << index[targ] << ") ";
      }
      std::cout << std::endl;
      //...
    }
    //...
  };

  out-edges: (0,1) (0,2) (0,3) (0,4)

The in_edges() function of the BidirectionalGraph interface provides access to all the in-edges of a vertex through in-edge iterators. The in_edges() function is only available for the adjacency_list if bidirectionalS is supplied for the Directed template parameter. There is an extra cost in space when bidirectionalS is specified instead of directedS.

  template <class Graph> struct exercise_vertex {
    //...
    void operator()(const Vertex& v) const
    {
      //...
      std::cout << "in-edges: ";
      typedef typename graph_traits<Graph> GraphTraits;
      typename GraphTraits::in_edge_iterator in_i, in_end;
      for (boost::tie(in_i, in_end) = in_edges(v,g);
           in_i != in_end; ++in_i) {
        e = *in_i;
        Vertex src = source(e, g), targ = target(e, g);
        std::cout << "(" << index[src] << "," << index[targ] << ") ";
      }
      std::cout << std::endl;
      //...
    }
    //...
  };

  in-edges: (2,0) (3,0) (4,0)

Adjacent Vertices

Given the out-edges of a vertex, the target vertices of these edges are adjacent to the source vertex. Sometimes an algorithm does not need to look at the edges of the graph and only cares about the vertices. Therefore the graph interface also includes the adjacent_vertices() function of the AdjacencyGraph interface which provides direct access to the adjacent vertices. This function returns a pair of adjacency iterators. Dereferencing an adjacency iterator gives a vertex descriptor for an adjacent vertex.

  template <class Graph> struct exercise_vertex {
    //...
    void operator()(Vertex v) const
    {
      //...
      std::cout << "adjacent vertices: ";
      typename graph_traits<Graph>::adjacency_iterator ai;
      typename graph_traits<Graph>::adjacency_iterator ai_end;
      for (boost::tie(ai, ai_end) = adjacent_vertices(v, g);
           ai != ai_end; ++ai)
        std::cout << index[*ai] <<  " ";
      std::cout << std::endl;
    }
    //...
  };

  adjacent vertices: 0 1

Adding Some Color to your Graph

BGL attempts to be as flexible as possible in terms of accommodating how properties are attached to a graph. For instance, a property such as edge weight may need to be used throughout a graph object's lifespan and therefore it would be convenient to have the graph object also manage the property storage. On the other hand, a property like vertex color may only be needed for the duration of a single algorithm, and it would be better to have the property stored separately from the graph object. The first kind of property is called an internally stored property while the second kind is called an externally stored property. BGL uses a uniform mechanism to access both kinds of properties inside its graph algorithms called the property map interface, described in Section Property Map Concepts. In addition, the PropertyGraph concept defines the interface for obtaining a property map object for an internally stored property.

The BGL adjacency_list class allows users to specify internally stored properties through plug-in template parameters of the graph class. How to do this is discussed in detail in Section Internal Properties. Externally stored properties can be created in many different ways, although they are ultimately passed as separate arguments to the graph algorithms. One straightforward way to store properties is to create an array indexed by vertex or edge index. In the adjacency_list with vecS specified for the VertexList template parameter, vertices are automatically assigned indices, which can be accessed via the property map for the vertex_index_t. Edges are not automatically assigned indices. However the property mechanism can be used to attach indices to the edges which can be used to index into other externally stored properties.

In the following example, we construct a graph and apply dijkstra_shortest_paths(). The complete source code for the example is in examples/dijkstra-example.cpp. Dijkstra's algorithm computes the shortest distance from the starting vertex to every other vertex in the graph.

Dijkstra's algorithm requires that a weight property is associated with each edge and a distance property with each vertex. Here we use an internal property for the weight and an external property for the distance. For the weight property we use the property class and specify int as the type used to represent weight values and edge_weight_t for the property tag (which is one of the BGL predefined property tags). The weight property is then used as a template argument for adjacency_list.

The listS and vecS types are selectors that determine the data structure used inside the adjacency_list (see Section Choosing the Edgelist and VertexList). The directedS type specifies that the graph should be directed (versus undirected). The following code shows the specification of the graph type and then the initialization of the graph. The edges and weights are passed to the graph constructor in the form of iterators (a pointer qualifies as a RandomAccessIterator).

  typedef adjacency_list<listS, vecS, directedS,
                         no_property, property<edge_weight_t, int> > Graph;
  typedef graph_traits<Graph>::vertex_descriptor Vertex;
  typedef std::pair<int,int> E;

  const int num_nodes = 5;
  E edges[] = { E(0,2),
                E(1,1), E(1,3), E(1,4),
                E(2,1), E(2,3),
                E(3,4),
                E(4,0), E(4,1) };
  int weights[] = { 1, 2, 1, 2, 7, 3, 1, 1, 1};

  Graph G(edges, edges + sizeof(edges) / sizeof(E), weights, num_nodes);

For the external distance property we will use a std::vector for storage. BGL algorithms treat random access iterators as property maps, so we can just pass the beginning iterator of the distance vector to Dijkstra's algorithm. Continuing the above example, the following code shows the creation of the distance vector, the call to Dijkstra's algorithm (implicitly using the internal edge weight property), and then the output of the results.

  // vector for storing distance property
  std::vector<int> d(num_vertices(G));

  // get the first vertex
  Vertex s = *(vertices(G).first);
  // invoke variant 2 of Dijkstra's algorithm
  dijkstra_shortest_paths(G, s, distance_map(&d[0]));

  std::cout << "distances from start vertex:" << std::endl;
  graph_traits<Graph>::vertex_iterator vi;
  for(vi = vertices(G).first; vi != vertices(G).second; ++vi)
    std::cout << "distance(" << index(*vi) << ") = "
              << d[*vi] << std::endl;
  std::cout << std::endl;

  distances from start vertex:
  distance(0) = 0
  distance(1) = 6
  distance(2) = 1
  distance(3) = 4
  distance(4) = 5

Extending Algorithms with Visitors

Often times an algorithm in a library almost does what you need, but not quite. For example, in the previous section we used Dijkstra's algorithm to calculate the shortest distances to each vertex, but perhaps we also wanted to record the tree of shortest paths. One way to do this is to record the predecessor (parent) for each node in the shortest-paths tree.

It would be nice if we could avoid rewriting Dijkstra's algorithm, and just add that little bit extra needed to record the predecessors [1]. In the STL, this kind of extensibility is provided by functors, which are optional parameters to each algorithm. In the BGL this role is fulfilled by visitors.

A visitor is like a functor, but instead of having just one "apply" method, it has several. Each of these methods get invoked at certain well-defined points within the algorithm. The visitor methods are explained in detail in Section Visitor Concepts. The BGL provides a number of visitors for some common tasks including a predecessor recording visitor. The user is encouraged to write his or her own visitors as a way of extending the BGL. Here we will take a quick look at the implementation and use of the predecessor recorder. Since we will be using the dijkstra_shortest_paths() algorithm, the visitor we create must be a Dijkstra Visitor.

The functionality of the record_predecessors visitor is separated into two parts. For the storage and access of the predecessor property, we will use a property map. The predecessor visitor will then only be responsible for what parent to record. To implement this, we create a record_predecessors class and template it on the predecessor property map PredecessorMap. Since this visitor will only be filling in one of the visitor methods, we will inherit from dijkstra_visitor which will provide empty methods for the rest. The constructor of the predecessor_recorder will take the property map object and save it away in a data member.

  template <class PredecessorMap>
  class record_predecessors : public dijkstra_visitor<>
  {
  public:
    record_predecessors(PredecessorMap p)
      : m_predecessor(p) { }

    template <class Edge, class Graph>
    void edge_relaxed(Edge e, Graph& g) {
      // set the parent of the target(e) to source(e)
      put(m_predecessor, target(e, g), source(e, g));
    }
  protected:
    PredecessorMap m_predecessor;
  };

The job of recording the predecessors is quite simple. When Dijkstra's algorithm relaxes an edge (potentially adding it to the shortest-paths tree) we record the source vertex as the predecessor of the target vertex. Later, if the edge is relaxed again the predecessor property will be overwritten by the new predecessor. Here we use the put() function associated with the property map to record the predecessor. The edge_filter of the visitor tells the algorithm when to invoke the explore() method. In this case we only want to be notified about edges in the shortest-paths tree so we specify tree_edge_tag.

As a finishing touch, we create a helper function to make it more convenient to create predecessor visitors. All BGL visitors have a helper function like this.

  template <class PredecessorMap>
  record_predecessors<PredecessorMap>
  make_predecessor_recorder(PredecessorMap p) {
    return record_predecessors<PredecessorMap>(p);
  }

We are now ready to use the record_predecessors in Dijkstra's algorithm. Luckily, BGL's Dijkstra's algorithm is already equipped to handle visitors, so we just pass in our new visitor. In this example we only need to use one visitor, but the BGL is also equipped to handle the use of multiple visitors in the same algorithm (see Section Visitor Concepts).

  using std::vector;
  using std::cout;
  using std::endl;
  vector<Vertex> p(num_vertices(G), graph_traits<G>::null_vertex()); //the predecessor array
  dijkstra_shortest_paths(G, s, distance_map(&d[0]).
                          visitor(make_predecessor_recorder(&p[0])));

  cout << "parents in the tree of shortest paths:" << endl;
  for(vi = vertices(G).first; vi != vertices(G).second; ++vi) {
    cout << "parent(" << *vi;
    if (p[*vi] == graph_traits<G>::null_vertex())
      cout << ") = no parent" << endl;
    else
      cout << ") = " << p[*vi] << endl;
  }

  parents in the tree of shortest paths:
  parent(0) = no parent
  parent(1) = 4
  parent(2) = 0
  parent(3) = 2
  parent(4) = 3

Notes