[Coin-discuss] New Mittelmann benchmarks on MILP and network LPs

[Sebastian Nowozin]

Hello,

there are new benchmarks on Hans Mittelmann's benchmark page, just a
few days old:

Large Network-LP benchmark with Clp 1.6 doing an excellent job against
Cplex and Mosek,
http://plato.asu.edu/ftp/network.html

MILP free solver codes, Cbc 2.1, Glpk 4.26, Symphony, SCIP, LP_Solve
http://plato.asu.edu/ftp/milpf.html

In the second benchmark I find two points interesting.

 1. Some problems are solved well by a simple branch and bound solver
(mark._4_0, misc07, nug08, 30_95_100 (Cbc failing), solved fast by
LP_Solve) but many times slower by Cbc.  This might be due to a lucky
branching decision early on, but it somehow sticks out of Cbc's
otherwise great results.

 2. Some problems (cap6000, swath2, neos1, neos20 to neos23, the m20*
problems) are solved both by Cbc 2.1 and SCIP 1.0 (with Clp as backend
solver, the SCIP_L tab in the benchmark table).  However, for the
problems I listed SCIP has a huge advantage.  These problems might be
an interesting starting point to look for effective cuts/heuristics
SCIP knows about and Cbc does not (yet), as the speedup can't be due
to the backend solver, which is the same for both.

Sebastian