[boost] [BGL] relax.hpp regression

From: Vladimir Prus (ghost_at_[hidden])
Date: 2005-06-17 08:59:23

Hi,
some time ago I've reported a problem with dag_shortest_paths that occurs when
I'm searching for the *longest* path by appropriately changing inf/compare
operators.

I've fixed that problem by change in dag_shortest_paths.hpp, but looks like 
it's back after changes to relax.hpp.

I attach my original report, which has a test program. When using version
1.21 of graph/relax.hpp, it produces the expected outcome '2'. When using 
version 1.22, I get 2147483647

Revision 1.22 is:

  revision 1.22
  date: 27/10/2004 18:38; author: dgregor

  Merge from graph_devel_1_33_0 branch

And the only commit on that branch does not clarify anything either:

   revision 1.21.8.1 
   date: 15/10/2004 18:54; author: dgregor

   Added several algorithms:
  - random graph layout
  - Fruchterman-Reingold layout
  - A* search
  - Floyd-Warshall all-pairs shortest paths

  Fixes for several algorithms:
  - Documentation for Johnson all-pairs shortest paths
  - Bellman-Ford single-source shortest paths

Any ideas?

- Volodya


attached mail follows:


Hello,
it appears that the dag_shortest_paths algorithm is broken. The attached 
program tries to find the longest path by changing inf/comparison operators. 
The graph is

0  ----->   1  ------>  2
                ^
                |
3 ----------/

The edge weights are 1 for  0->1 and 1->2 and 2 for 3->1 and I'm searching for 
the longest path starting from '0'. The final result for '2' is 0x7FFFFFFF, 
while I'd expect to get '2'.

The problem is two-fold. First, the algorithm runs topological sort on the 
graph, and process the vertices in topological order -- which include the '3' 
vertex. That vertex is not reachable from '0', so it should not be included 
in the search at all.

The second problem is that the distance to that vertex is set to -inf. When 
combining that distance with edge weight, due to a bug we get +inf. Here's 
the code:

    template <class T>
    struct closed_plus
    {
      // std::abs just isn't portable :(
      template <class X>
      inline X my_abs(const X& x) const { return x < 0 ? -x : x; }

      T operator()(const T& a, const T& b) const {
        using namespace std;
        T inf = (numeric_limits<T>::max)();
        if (b > 0 && my_abs(inf - a) < b)
          return inf;
        return a + b;
      }
    };

When 'a' is -inf (as is for '3') inf - a overflows and gives -1, and when b is 
greater than 1, "inf" is returned.

I think that the right solution is to change the algorithm so that it does not 
traverse the vertices not reachable from the start one. Besides fixing my 
problem, this will also speed up the algorithm. For example, if you have a 
forest and call dag_shortest_path on the root of just one tree, the algorithm 
will travese all trees, setting pretty useless values for all of them.

I would like to commit the attached patch. All tests still pass on gcc when 
it's applied. Comments? 

- Volodya