Are our random optimization methods better than resourcematched deterministic hillclimbing?
For our deterministic entry, starting with randomized configurations,
using a null entry tabledriven scheme, we took the best of five
iterated nextdescent hillclimbing rounds (approximately 40 million
total attempts), in both random (mr) and priority order (m) as
described in Section 3.3. The random and priority orders
of iteration did not register any difference.
Pairwise comparisons consistently showed the iterated nextdescent
hillclimbing rounds were different than all the randomized methods
tested, except the unminimized tabledriven schemes (sa1 &
sigmet1) in all terms, minimized tabledriven schemes (sa1/m &
sigmet1/m) in 1999, and the unminimized acceptbased adaptive
scheme (sa3) in Fall 2000.
On average, the iterated nextdescent hillclimbing scores (m & mr)
were an order of magnitude greater than all the
others^{}, with
wider standard deviations (Tables 4.3 through
4.5), forcing us to conclude our randomized methods
have a distinct advantage over simple hillclimbing.

