Andrew,
My student and I noticed some incorrect behavior (in our opinion) for
the time limit parameter. When solving a MILP, the time limit is only
checked while solving the root LP relaxation and then before solving
each node in the b&b tree.
We came across an odd case where one of the nodes takes a long time to
solve (maybe it's even a numerical issue). At any rate, the time limit
does not work correctly, because it isn't checked until after the node
is solved.
A quick fix is to add a single line at around line 736 of glpmip1.c in
function mip_solve_node():
lpx_set_real_parm(lp, LPX_K_TMLIM, tree->tm_lim + tree->tm_beg -
utime());
This just passes a time limit to the LP algorithm. We tried this and it
seems to work OK, though we haven't yet recreated the real problem
situation.
This solution isn't quite correct, because some time passes between the
last check and this check, so the value above could actually be
negative. If that were the case, it wouldn't work correctly. Better
would be something like:
double remaining_time;
remaining_time = tree->tm_beg + tree->tm_lim - utime();
/* Check that there is always a small positive limit */
if (tree->tm_lim > 0 && remaining_time <= 0.0)
remaining_time = 0.1;
lpx_set_real_parm(lp, LPX_K_TMLIM, remaining_time);
I haven't tested that, though.
Brady
--
Brady Hunsaker
Assistant Professor
Industrial Engineering
University of Pittsburgh
http://www.engr.pitt.edu/hunsaker/
_______________________________________________
Bug-glpk mailing list
Bug-glpk@gnu.org
http://lists.gnu.org/mailman/listinfo/bug-glpk
