Tablebase test: SF18, Reckless and PlentyChess

[Jouni]

Code: Select all

Score of PlentyChess TB vs PlentyChess: 321 - 259 - 420 [0.531]
...      PlentyChess TB playing White: 159 - 130 - 211  [0.529] 500
...      PlentyChess TB playing Black: 162 - 129 - 209  [0.533] 500
...      White vs Black: 288 - 292 - 420  [0.498] 1000
Elo difference: 21.6 +/- 16.4, LOS: 99.5 %, DrawRatio: 42.0 %
1000 of 1000 games finished.

Score of Reckless TB vs Reckless: 321 - 271 - 408 [0.525]
...      Reckless TB playing White: 155 - 136 - 209  [0.519] 500
...      Reckless TB playing Black: 166 - 135 - 199  [0.531] 500
...      White vs Black: 290 - 302 - 408  [0.494] 1000
Elo difference: 17.4 +/- 16.6, LOS: 98.0 %, DrawRatio: 40.8 %
1000 of 1000 games finished.

Score of Stockfish18 TB vs Stockfish18: 269 - 264 - 467 [0.502]
...      Stockfish18 TB playing White: 140 - 123 - 237  [0.517] 500
...      Stockfish18 TB playing Black: 129 - 141 - 230  [0.488] 500
...      White vs Black: 281 - 252 - 467  [0.514] 1000
Elo difference: 1.7 +/- 15.7, LOS: 58.6 %, DrawRatio: 46.7 %
1000 of 1000 games finished.

Conditions: 6 piece tablebases in SSD, 10+0.1 level and endgames_cdb95105.epd book. Even with 1000 games we see SF don't gain from 6 piece tbs. And with 7 piece it has time losses in TCEC .

[AndrewGrant]

The time loss comment might be a bit misleading. I am not suggesting you are trying to mislead, to be clear. Just adding a note for any other readers.

SF is not losing on time because of the cost of 7-man probes in Search. It was losing on time because it was trying to provide better PVs after the fact based on the probes. Something that exists purely cosmetically, to help the user.

At least, this was my understanding. I don't follow TCEC too closely in recent years.