What are the best engines against humans in chess?

[Uri Blass]

Even if all humans lose it is possible to decide that the engine who get winning position in smaller average number of moves is the best.

We cannot be sure if a position is winning but we can dedice that if the best engine can beat itself then the position is probably winning.

I also wonder what is the average number of moves that engines need to get a winning position with white or with black against different level of rating of humans.

[Dann Corbit]

It might be possible to mine a statistical answer from the LiChess games.

[jefk]

well indeed from lichess (or new separate tests,) some answers could be derived although
there not so many bots (yet) i think like P4, RebelEas, and Sf full strength.

imo it depends on the rating level of the humans, below master level P4 with
eg. my gambit book could give faster mates on average than SF.
Above master level, RebelEas may still do a (very) good job. And it also
depends on the time frame, ofcourse; in bullet, even Nakamura
would stand much chance against P4 or -certainly- RebelEas; we've
seen how he played against leela-knight odds (yes annoying for a human
and most likely these super'agressive' (sharp) engines would also be
quite annoying at fast time frames (bullet or fast blitz eg 3-0).

Only at top-Gm level with slower time controls, eg rapid or classical, SF may
perform 'better' (faster mates), especially if such a top GM would have
prepared against the engine (and it's book). my 2 cnts (but a real world
experiment could confirm it; if there would be any volunteers..

[Hai]

Even if all humans lose it is possible to decide that the engine who get winning position in smaller average number of moves is the best.

We cannot be sure if a position is winning but we can dedice that if the best engine can beat itself then the position is probably winning.

I also wonder what is the average number of moves that engines need to get a winning position with white or with black against different level of rating of humans.

https://en.chessbase.com/post/brand-new ... e-fritz-20

[jefk]

yep F20 at lower elo (but also other weak engines with human style) could be used to simulate humans at
various levels (and with various style) and in a weakened F20-engine vs P4 (or R-Ea or SF) engine series
of games observe the results, and draw conclusions.

It probably also would result in many interesting games.

[Hai]

yep F20 at lower elo (but also other weak engines with human style) could be used to simulate humans at
various levels (and with various style) and in a weakened F20-engine vs P4 (or R-Ea or SF) engine series
of games observe the results, and draw conclusions.

It probably also would result in many interesting games.

Lower elo means??
Is it possible to play against Tal or Fischer at full strength or even stronger? or must we play against a 1500 elo Tal or Fischer?

[chrisw]

One could well posit that given the current search depths or more that an engine can achieve, plus the current huge win rates against top humans, that, given similar search characteristics, a maniac engine (eg sacrificial king attacker) will bring down the human player faster and with more style. If so, this would upend the paradigm of high accuracy evals.

[jefk]

hai wrote

lower elo means

well i was talking about using engines (F20 or whatever) to simulate a human (*)
to play against an aggressive engine at their max strength (to answer the
question by Uri B)

(*) thus at weakened strength (elo) ie -much- lower strength elo) than usual;
apparently with F 20 you can also let those old gm's play at lower elo but i doubt
if this would give realistic club player style, maybe (if a certain person really tries
to emulate such a GM), maybe, probably not.

Maybe by mentioning F20 you meant it could also be an aggressive engine (eg in Tal style)
at highest Elo, thus possibly winning faster than SF against most humans. Fine with me.
Why an aggressive style (to win fast) ? Well Tal (and Morphy) made mistakes, but take
their style, or more 'aggressive' like P4, without the blunders and then such an engine
wins faster than eg. a Capablanca or a Karpov style.

cw wrote

a maniac engine (eg sacrificial king attacker) will bring down the human player faster and with more style.
If so, this would upend the paradigm of high accuracy evals.

yep, but then the accuracy evals of eg. a Rebel-Eas also are not too bad i think;
although not optimized for 'accuracy' (ie max Elo in engine vs engine games). And maybe
even P4 (or later P5) could do the job at least against humans below IM level

[Rebel]

The current best king attackers are :

Code: Select all

PGN database      : pgn\ka.pgn

Measure King Attacks : 83.900 games

Engine                  Won Loss   Perc   ATTACK
CSTal 2.1 EAS           1693  8156  17.2% 105676
Patricia 3.0 avx2        548 10487   5.0%  97395
Patricia 3.1 avx2        589 10460   5.3%  96068
Patricia 241106 avx2    1434  8232  14.8%  77902
Rebel Extreme avx2      2014  6490  23.7%  66522
Velvet 8.1.0 risky      1784  6761  20.9%  45813
Rebel EAS 2.0 avx2      4702  2886  62.0%  38882
Stockfish 231014 avx2    547     3  99.5%  38589
Stockfish 230911 avx2    547     6  98.9%  38566
Stockfish 231116 avx2    564     6  98.9%  37683
Stockfish 230929 avx2    559     5  99.1%  37369
Rebel EAS avx2          5630 12052  31.8%  36785
Stockfish 231024 avx2    559     4  99.3%  36223
Stockfish 231107 avx2    557     6  98.9%  35492

Not included yet is the new Patricia.

The other top engines, all games are from Stefan's rating list :

Code: Select all

 PGN database      : pgn\uho.pgn

 Measure King Attacks : 60.500 games

 Engine                  Won Loss   Perc   ATTACK
 Stockfish 250427 a512   6822  1000  87.2%  39131
 Stockfish 17.1 250330   6744  1006  87.0%  38124
 Horsie 1.0 a512         2274  5246  30.2%  34236
 KomodoDragon 3.3 avx2   3535  3948  47.2%  32883
 Torch 3.1 a512          5633  2092  72.9%  31147
 Ethereal 14.38 avx2     2421  5069  32.3%  30987
 Integral 7.0 a512       3451  4249  44.8%  30851
 Clover 8.1 avx2         1957  5423  26.5%  29873
 RubiChess 240817 avx2   2151  5452  28.3%  29840
 Alexandria 8.0 a512     4019  3497  53.5%  28426
 Starzix 6.0 a512        2100  5506  27.6%  27646
 PlentyChess 5.0 a512    4273  3380  55.8%  27419
 Berserk 250307 a512     4225  3282  56.3%  27144
 Obsidian 15.14 a512     5061  2486  67.1%  26705

[Peter Berger]

Even if all humans lose it is possible to decide that the engine who get winning position in smaller average number of moves is the best.

We cannot be sure if a position is winning but we can dedice that if the best engine can beat itself then the position is probably winning.

I also wonder what is the average number of moves that engines need to get a winning position with white or with black against different level of rating of humans.

If top humans played in the usual ambitious way they use to play against each other, I don't think there would be a very noteable difference between various engines performance-wise or if there were one it would be mostly coincidence.

The real question is how this would look if humans tried to play a bit like "Father", without any other ambition than to draw games and/or drag them on.

I think top human players would even get a pretty decent number of draws against Stockfish this way, the same would be much more difficult against Chess System Tal EAS.