TODO

* use boost::sub_range
* add example to metrics

* euclidean metric shouldn't be parametrized

* doxygen comment in code should start from big letter

* check boost concepts

* separate in/out, type, name, description in doxygen

* disable typedefs warning in doxygen
* knapsack unbounded fptas is not realy fptas
  there is a problem with number of objects.
  The number of objects is included in the multiplier.
  In the current version we use incorrect number of items.
  One of the solution could be to group the small eelements in the bigger bundles.
* doxygen does not see some of the references
    (http://siekiera.mimuw.edu.pl/~paal/stein.html -> paal::data_structures::Voronoi.)
* alternative kanpsack algorithm (n*(1/epsilon)^2)
* make better traits, maybe it supposed to look like std::iterator traits
* function comparing 2 metrics

* ITERATIVE ROUNDING
** error checking for generalized assignment
** move and copy constructor in glp class
** glp.get_columns() returns columns in random order
** LP formulation in separate dir
** improve row generation ?
** resolving solution by moving a point ?
** cplex support ???

* lp
** (glpk) in the resolve we don't allways want to call
   glp_adv_basis, but run it only when glp_simplex fails.
   This is not done beacause of the bug in the glpk.
   It should be done when the bug is fixed.
   http://lists.gnu.org/archive/html/bug-glpk/2014-06/msg00000.html


* GREEDY
** steiner forest (conjecture)

* make install

* LSH regression
** computation of params and using them by default:
*** w_param and hash_functions_per_point: by optimizing probability of hash
    collisions
** stream version of algorithm
*** correct computing max point dimensions in lsh_nearest_neighbors_regression
    and all tests
** classification version
** multithreading
** convenient binary