[RFC] Prefer suboptimal moves to time losses

[robertnurnberg]

This PR is a follow up to #6422. We now also abort DTZ probes at root if we run low on time.

Edit: Apologies about previously published numbers, they were for an old master binary without the pv extension dtz patch. Corrected numbers show no measurable impact at the tested TC.

> ./fastchess -engine name="patch" cmd=../stockfish/src/stockfish.patch -engine name="master" cmd=../stockfish/src/stockfish.master -each tc=1+0.05 option.SyzygyPath=/disk1/syzygy/3-4-5-6/WDL:/disk2/syzygy/3-4-5-6/DTZ -openings file=endgames.epd format=epd -rounds 500 -repeat -concurrency
...
Results of patch vs master (1+0.05, 1t, 16MB, endgames.epd):
Elo: -6.60 +/- 10.16, nElo: -14.00 +/- 21.53
LOS: 10.13 %, DrawRatio: 71.00 %, PairsRatio: 0.81
Games: 1000, Wins: 170, Losses: 189, Draws: 641, Points: 490.5 (49.05 %)
Ptnml(0-2): [15, 65, 355, 54, 11], WL/DD Ratio: 0.36
--------------------------------------------------

Player: master
  Timeouts: 22
  Crashed: 0
Player: patch
  Timeouts: 27
  Crashed: 0

Finished match

No functional change.

[robertnurnberg]

Pushed a change that always uses WDL as fall-back at root.

[robertnurnberg]

Some comments from my side: firstly, I would completely understand if maintainers say they would rather not open this can of worms and leave master as it is. Secondly, if this gets merged it should be on the understanding that the code is off-limits for any tinkering with constants and so on. We should come to some agreement on what approach makes sense, and then stick with that.

If we want to go ahead with this, then imo the main thing to discuss is the timeBudget allowed for the DTZ probes at root. Is 1% of the remaining time enough for the conditions at TCEC, for example, to guarantee all root moves can be DTZ probed? (The standard increment at TCEC is 3s, so 1% would be 30ms.)

[vondele]

I think it is a bit a can of worms... my first question would be how does this work if we're not playing with TC, but fixed depth or nodes?

[robertnurnberg]

Pushed a fix for nodestime option (also for the pv extension part). Also corrected small test numbers, sorry about previously misleading numbers.

[robertnurnberg]

Some more results on the standard book, at different TC, with my setup. Just for completeness.

Results of patch vs master (1+0.01, 1t, 16MB, UHO_Lichess_4852_v1.epd):
Elo: -0.42 +/- 4.63, nElo: -0.61 +/- 6.81
LOS: 42.99 %, DrawRatio: 41.62 %, PairsRatio: 0.99
Games: 10000, Wins: 3002, Losses: 3014, Draws: 3984, Points: 4994.0 (49.94 %)
Ptnml(0-2): [280, 1189, 2081, 1163, 287], WL/DD Ratio: 1.55
--------------------------------------------------

Player: patch
  Timeouts: 272
  Crashed: 0
Player: master
  Timeouts: 292
  Crashed: 0

Results of patch vs master (1, 1t, 16MB, UHO_Lichess_4852_v1.epd):
Elo: 2.47 +/- 5.14, nElo: 3.27 +/- 6.81
LOS: 82.66 %, DrawRatio: 38.78 %, PairsRatio: 1.03
Games: 10000, Wins: 3430, Losses: 3359, Draws: 3211, Points: 5035.5 (50.36 %)
Ptnml(0-2): [425, 1084, 1939, 1099, 453], WL/DD Ratio: 2.77
--------------------------------------------------

Player: master
  Timeouts: 298
  Crashed: 0
Player: patch
  Timeouts: 278
  Crashed: 0

Results of patch vs master (5, 1t, 16MB, UHO_Lichess_4852_v1.epd):
Elo: 2.15 +/- 3.94, nElo: 3.72 +/- 6.81
LOS: 85.80 %, DrawRatio: 48.18 %, PairsRatio: 1.04
Games: 10000, Wins: 2774, Losses: 2712, Draws: 4514, Points: 5031.0 (50.31 %)
Ptnml(0-2): [122, 1147, 2409, 1191, 131], WL/DD Ratio: 1.21
--------------------------------------------------

Player: master
  Timeouts: 110
  Crashed: 0
Player: patch
  Timeouts: 104
  Crashed: 0

Results of patch vs master (10+0.1, 1t, 16MB, UHO_Lichess_4852_v1.epd):
Elo: -2.92 +/- 3.58, nElo: -5.54 +/- 6.81
LOS: 5.53 %, DrawRatio: 51.28 %, PairsRatio: 0.93
Games: 10000, Wins: 2583, Losses: 2667, Draws: 4750, Points: 4958.0 (49.58 %)
Ptnml(0-2): [55, 1206, 2564, 1118, 57], WL/DD Ratio: 1.11
--------------------------------------------------

Player: patch
  Timeouts: 37
  Crashed: 0
Player: master
  Timeouts: 42
  Crashed: 0