Official Release of OliThink 5.7.5 including a Java-GUI

[OliverBr]

5.8.1 has been released with another kind of pruning (SEE pruning). I didn't like all this pruning for a long time, but it is, actually, working.

I would like to show a match against Fruit 2.1, which ended about move 32 (black resigning), even though OliThink opened with 1. f4.

[pgn][Event "Computer Chess Game"]
[Site "Olivers-MacBook-2.local"]
[Date "2020.09.24"]
[Round "-"]
[White "OliThink 5.8.1"]
[Black "Fruit 2.1"]
[Result "1-0"]
[TimeControl "40/60"]
[Annotator "2. -0.02 1... +0.11"]

1. f4 Nf6 {+0.11/13 3} 2. Nf3 {-0.02/16 1.0} e6 {+0.07/12 1.3} 3. Nc3
{+0.02/16 1.4} d5 {+0.07/12 2.6} 4. e3 Bb4 {+0.03/11 1.8} 5. a3
{+0.19/17 1.1} Bd6 {+0.05/11 1.5} 6. b3 {+0.16/15 1.2} O-O {+0.05/10 1.4}
7. Bb2 {+0.15/17 1.0} Nc6 {+0.14/11 1.7} 8. Nb5 {+0.25/19 2.3} Be7
{+0.00/11 1.6} 9. Be2 {+0.22/20 0.2} a6 {+0.13/11 1.5} 10. Nbd4
{+0.12/20 2.1} Nxd4 {+0.11/12} 11. Bxd4 {+0.09/20 1.3} Ne4 {+0.12/12 0.4}
12. O-O {+0.15/18 1.2} c5 {+0.05/12} 13. Bb2 {+0.11/17 2.3} f6 {+0.08/12}
14. a4 {+0.20/17 1.4} Bd7 {+0.03/12 2.2} 15. a5 {+0.20/18 0.2} Qc7
{+0.00/12 2.5} 16. Kh1 {+0.12/17 1.7} Kh8 {+0.00/11 3} 17. Qe1 c4
{+0.03/10 1.4} 18. bxc4 {+0.25/17 1.1} Bb4 {-0.11/10 0.3} 19. cxd5
{+0.31/16 1.0} exd5 {-0.03/10 1.8} 20. c3 {+0.23/18 0.2} Bxa5
{-0.02/11 1.5} 21. Qb1 {+0.23/17 1.5} Rfd8 {+0.00/9 1.5} 22. Qa2
{+0.19/19 1.8} b6 {-0.01/11} 23. Ba3 {+0.17/18 1.3} Bg4 {-0.04/11 0.5} 24.
Bb4 {+0.21/18 0.9} Bxb4 {+0.02/10 1.8} 25. cxb4 {+0.12/18 0.6} a5
{+0.08/10 0.7} 26. Rfc1 {+0.16/15 1.4} Qe7 {+0.20/11} 27. bxa5
{-0.21/16 1.4} d4 {+0.06/10 0.6} 28. Nxd4 {+0.12/19 3} Rxa5 {-0.22/11} 29.
Qb2 {+1.84/19 1.4} Rxa1 {-1.52/11 8} 30. Rxa1 {+1.74/18 1.7} Bxe2
{-0.77/12 1.7} 31. Nc6 {+2.80/18 2.0} Qe8 {-2.83/13 5} 32. Nxd8 h5
{-2.92/10 0.5}
{Black resigns} 1-0
[/pgn]

[OliverBr]

Here is a tournament:

Code: Select all

40/30:
   # PLAYER                   :  RATING  ERROR  POINTS  PLAYED   (%)     W    D     L  D(%)  CFS(%)
   1 Glaurung 2.2             :     104     20   644.5    1000  64.5   561  167   272  16.7     100
   2 OliThink 5.8.1           :       0   ----  1915.5    4000  47.9  1527  777  1696  19.4      82
   3 GreKo 2020.03            :      -9     20   486.5    1000  48.6   394  185   421  18.5      60
   4 Scorpio_2.8.8 MCTS+NN    :     -13     19   481.5    1000  48.1   409  145   446  14.5      68
   5 WyldChess                :     -20     19   472.0    1000  47.2   332  280   388  28.0     ---

White advantage = 6.00 +/- 5.09
Draw rate (equal opponents) = 19.71 % +/- 0.64

OliThink does not have any chess knowledge other than mobility. "Double pawns", "King safety", "King shelter", "Dominate center", "Rook on 7th file", "Bishop pair", "Develop knights/bishops first", etc... are unknown.

[Alayan]

5.8.0 has been released. Here a small tournament with famous names, but ultra quick time control:

Code: Select all

   # PLAYER                     :  RATING  ERROR  POINTS  PLAYED   (%)     W     D     L  D(%)  CFS(%)
   1 Stockfish 11 64            :     752     16  9004.0   10007  90.0  8410  1188   409  11.9     100
   2 Ethereal 12.55 (POPCNT)    :     621     14  7912.0   10005  79.1  7213  1398  1394  14.0     100
   3 Minic 2.16                 :     183     10  3606.0   10005  36.0  2942  1328  5735  13.3     100
   4 Glaurung 2.2               :     119     10  2890.0   10005  28.9  2306  1168  6531  11.7     100
   5 OliThink 5.8.0             :       0   ----  1603.0   10008  16.0  1159   888  7961   8.9     ---

White advantage = 27.74 +/- 2.92
Draw rate (equal opponents) = 23.37 % +/- 0.40

Ethereal was a positive surprise, while Minic a negative one, I would have guessed Ehereal about 520 and Minic about 280

What was the exact TC you used ? Base+increment.

Stockfish 11 is known to regress vs Stockfish 10 at ultra-fast bullet (significantly below 10s+0.1s) because of a search tune that scales much better to longer time controls.

While on the other hand, in Ethereal 12.54 I fixed what was arguably a time management bug. Ethereal would use its time as if it had no move overhead, then play as fast as possible once below the move overhead threshold. Fixing this gained dozens of elo at sudden death and significant elo at hyper-fast bullet. The default move overhead is also back to 100ms after a year at 250ms, if the time control is really short this might make a difference.

[OliverBr]

What was the exact TC you used ? Base+increment.

I am sorry for the late answer: It was 40/2, which is 40 moves in 2 seconds. I would call this ultra-sonic-bullet.

[OliverBr]

OliThink 5.8.3 has been released.
Still, there is no chess knowledge other than mobility. There are no square tables, no tuning of numbers, no use of external data like Syzygy.

ELO gain since 5.8.1: 58
ELO gain since 5.7.5: 114

These gauntlet tournaments may illustrate the improvement.

Code: Select all

   # PLAYER             :  RATING  ERROR  POINTS  PLAYED   (%)     W    D     L  D(%)  CFS(%)
   1 Weiss 1.0          :     151     18   984.0    1400  70.3   845  278   277  19.9      94
   2 Murka 3 x64        :     132     17   952.5    1400  68.0   813  279   308  19.9     100
   3 Glaurung 2.2       :      57     17   813.0    1400  58.1   671  284   445  20.3     100
   4 OliThink 5.8.3     :       0   ----  1450.5    4200  34.5  1030  841  2329  20.0     ---

compared to 5.8.1:

Code: Select all

   # PLAYER            :  RATING  ERROR  POINTS  PLAYED   (%)    W    D     L  D(%)  CFS(%)
   1 Weiss 1.0         :     210     19  1075.0    1400  76.8  949  252   199  18.0      86
   2 Murka 3 x64       :     196     17  1055.0    1400  75.4  936  238   226  17.0     100
   3 Glaurung 2.2      :     110     17   912.0    1400  65.1  787  250   363  17.9     100
   4 OliThink 5.8.1    :       0   ----  1158.0    4200  27.6  788  740  2672  17.6     ---

and to 5.7.5:

Code: Select all

   # PLAYER            :  RATING  ERROR  POINTS  PLAYED   (%)     W    D     L  D(%)  CFS(%)
   1 Weiss 1.0         :     261     20  1142.0    1400  81.6  1023  238   139  17.0      76
   2 Murka 3 x64       :     251     19  1130.0    1400  80.7  1020  220   160  15.7     100
   3 Glaurung 2.2      :     172     19  1018.5    1400  72.8   896  245   259  17.5     100
   4 OliThink 5.7.5    :       0   ----   909.5    4200  21.7   558  703  2939  16.7     ---

[Graham Banks]

So, in CCRL 40/15 terms, what would you estimate its Elo to be?

[OliverBr]

So, in CCRL 40/15 terms, what would you estimate its Elo to be?

It's difficult to say, because my the tests are only playing blitz games. E.g. in 40/2 we can safely assume greater than 2700.

I could start a little tourney 40/15 against 2 or 3 other engines (need linux binary or source). What engines do you propose?

[Graham Banks]

So, in CCRL 40/15 terms, what would you estimate its Elo to be?

It's difficult to say, because my the tests are only playing blitz games. E.g. in 40/2 we can safely assume greater than 2700.

I could start a little tourney 40/15 against 2 or 3 other engines (need linux binary or source). What engines do you propose?

The same ones are fine.

[mvanthoor]

OliThink does not have any chess knowledge other than mobility. "Double pawns", "King safety", "King shelter", "Dominate center", "Rook on 7th file", "Bishop pair", "Develop knights/bishops first", etc... are unknown.

Have you ever thought about adding this? Your engine is already at +/- 2500 Elo without this knowledge. Stuff like passed pawn evaluation is known to add over 100 Elo in some engines. According to MadChess by Eric Madsen:

https://www.madchess.net/

+ 119 for passed pawns.
+ 63 for king safety.

OliThink 5.6.9 64-bit rates 2531 at CCRL 40/15.

You said you never achieved to get your engine up to 2700, in a different thread. If Eric's Elo increases also hold for Olithink, adding Passed pawns and king safety evaluation could net you 182 Elo, for a total of 2713. (And your current version will probably already be stronger than 2531. As it scores -172 against Glaurung, estimated strength would be around 2723. If you could add 182 Elo to that, you'd hit 2900...)

[OliverBr]

The same ones are fine.

Very well.. I have started a gauntlet tournament, 40 rounds à 2games (repeat opening), Time Control 40/15 (or 40/900 in seconds). In a couple of hours we will have a result

Code: Select all

./cutechess-cli \
-engine cmd=olithink583 proto=xboard \
-engine cmd=./glaurung proto=uci option.Threads=1 \
-engine cmd=weiss proto=uci \
-engine cmd=murka3 proto=uci \
-each tc=40/900 dir=bin \
-tournament gauntlet \
-games 2 -rounds 40 -concurrency 32 \
-openings file=openings/book_3moves_cp1-24_10944pos.pgn -repeat \
-pgnout tourn.pgn min

EDIT: 26 minutes later, not one of the 32 games has been finished yet .) This is really a marathon.