Using a SPSA tuning session to optimize the time management
parameters.
With SPSA tuning it is not always possible to say where improvements
came from. Maybe some variables changed randomly or because result
was not sensitive enough to them. So my explanation of changes will
not be necessarily correct, but here it is.
• When decrease of thinking time was added by Joost a few months ago
if best move has not changed for several plies, one more competing
indicator was introduced for the same purpose along with increase
in score and absence of fail low at root. It seems that tuning put
relatively more importance on that new indicator what allowed to save
time.
• Some of this saved time is distributed proportionally between all
moves and some more time were given to moves when score dropped a lot
or best move changed.
• It looks also that SPSA redistributed more time from the beginning to
later stages of game via other changes in variables - maybe because
contempt made game to last longer or for whatever reason.
All of this is just small tweaks here and there (a few percentages changes).
STC (10+0.1):
LLR: 2.96 (-2.94,2.94) [0.00,4.00]
Total: 18970 W: 4268 L: 4029 D: 10673
http://tests.stockfishchess.org/tests/view/5a9291a40ebc590297cc8881
LTC (60+0.6):
LLR: 2.95 (-2.94,2.94) [0.00,4.00]
Total: 72027 W: 12263 L: 11878 D: 47886
http://tests.stockfishchess.org/tests/view/5a92d7510ebc590297cc88ef
Additional non-regression tests at other time controls
Sudden death 60s:
LLR: 2.95 (-2.94,2.94) [-4.00,0.00]
Total: 14444 W: 2715 L: 2608 D: 9121
http://tests.stockfishchess.org/tests/view/5a9445850ebc590297cc8a65
40 moves repeating at LTC:
LLR: 2.95 (-2.94,2.94) [-4.00,0.00]
Total: 10309 W: 1880 L: 1759 D: 6670
http://tests.stockfishchess.org/tests/view/5a9566ec0ebc590297cc8be1
This is a functional patch only for time management, but the bench
does not reflect this because it uses fixed depth search, so the number
of nodes does not change during bench.
No functional change.
This patch simplifies the time management code, removing the extra
thinking time for moves with draw PV and increasing thinking time
for all moves proportionally by around 4%.
Last time when the time management was carefully tuned was 1.5-2 years
ago. As new patches were getting added, time management was drifting out
of optimum. This happens because when search becomes more precise pv and
score are becoming more stable, there are less fail lows, best move is
picked earlier and there are less best move changes. All this factors are
entering in time management, and average time per move is decreasing with
more and more good patches. For individual patches such effect is small
(except some) and may be up or down, but when there are many of them,
effect is more substantial. The same way benchmark with more and more
patches is slowly drifting down on average.
So my understanding that back in October adding more think time for draw
PV showed positive Elo because time management was not well tuned, there
was more time available, and think_hard patch applied this additional time
to moves with draw PV, while just retuning back to optimum would recover Elo
anyway. It is possible that absence of contempt also helped, as SF9 is showing
less 0.0 scores than the October version.
Anyway, to me it seems that proper place to deal with draw PV is search, and
contempt sounds as much better solution. In time management there is little
additional elo, and if some code is not helping like removed here, it is better
to discard it. It is simpler to find genuine improvement if code is clean.
• Passed STC:
LLR: 2.95 (-2.94,2.94) [-3.00,1.00]
Total: 20487 W: 4558 L: 4434 D: 11495
http://tests.stockfishchess.org/tests/view/5a7706ec0ebc5902971a9854
• Passed LTC:
LLR: 2.96 (-2.94,2.94) [-3.00,1.00]
Total: 41960 W: 7145 L: 7058 D: 27757
http://tests.stockfishchess.org/tests/view/5a778c830ebc5902971a9895
• Passed an additional non-regression [-5..0] test at the time control
of 60sec for the game (sudden death) with disabled draw adjudication:
LLR: 2.95 (-2.94,2.94) [-5.00,0.00]
Total: 8438 W: 1675 L: 1586 D: 5177
http://tests.stockfishchess.org/tests/view/5a7c3d8d0ebc5902971a9ac0
• Passed an additional non-regression [-5..0] test at the time control
of 1sec+1sec per move with disabled draw adjudication:
LLR: 2.97 (-2.94,2.94) [-5.00,0.00]
Total: 27664 W: 5575 L: 5574 D: 16515
http://tests.stockfishchess.org/tests/view/5a7c3e820ebc5902971a9ac3
This is a functional change for the time management code.
Bench: 4983414
Simplify time management code by removing hard stops for unchanging first root moves.
Search is now stopped earlier at the end iteration if it did not have fail-lows at root.
This simplification also fixes pondering bug. Ponder flag was true by default
and cutechess-cli doesn't change it to false even though no pondering is possible.
Fix the issue by setting the default value of 'Ponder' flag to false.
10+0.1:
ELO: 3.51 +-3.0 (95%) LOS: 99.0%
Total: 20000 W: 3898 L: 3696 D: 12406
40+0.4:
ELO: 1.39 +-2.7 (95%) LOS: 84.7%
Total: 20000 W: 3104 L: 3024 D: 13872
60+0.06:
LLR: 2.95 (-2.94,2.94) [-3.00,1.00]
Total: 37231 W: 5333 L: 5236 D: 26662
Stopped run at 100+1:
LLR: 1.09 (-2.94,2.94) [-3.00,1.00]
Total: 37253 W: 4862 L: 4856 D: 27535
Resolves#523Fixes#510
This is a very discussed patch with many
argumentations pro and against. The fact is
it passed both STC:
LLR: 2.96 (-2.94,2.94) [-1.50,4.50]
Total: 16305 W: 3001 L: 2855 D: 10449
And LTC
LLR: 2.95 (-2.94,2.94) [0.00,6.00]
Total: 34273 W: 5180 L: 4931 D: 24162
Although it is true that a correct test should
include foreign engines, we commit it anyhow so
people can test it out in the wild, under broader
conditions.
bench: 7384368
