Perfect play

[mmt]

Let's take the programs and the hardware as in the chess.com computer chess championship - basically a $15k workstation for both LC0 and SF11. What's your guess at the win % when playing black that a 32-piece TB would achieve against them? Anybody thinks 100%? Or are we getting close to the top possible ELO? A little complication is that varying from the perfect line might achieve a higher score vs current programs.

Could we approximate the top possible ELO by playing these engines at higher time controls vs classical time controls? We'd see what type of graph we'd get - is it maybe a logistic function or a log function at this current top level? We could start at shorter time controls as this would take forever.

[Spliffjiffer]

if SF plays 1.e4 and as black the 32 TB-entity sees that every single respond ends in a draw, how would this entity recognize in which of these lines SF will blunder the game away...the entity eighther must be god or a heuristic algorythm to evaluate the difficulty of the remaining positions

[Ovyron]

Unassisted engines are still far away from top correspondence players. Instead of trying to give both engines the fastest hardware available it'd be more fruitful to get in contact with such a player and ask them if they believe they could beat the unassisted engine (if you can't find such a player then I concede this is a lost case.)

Because rarely an engine will change its mind from the move it first picked at top hardware at 12 hour/move, and the times it changes their mind, it may not matter, while a skilled centaur at a week/move could prepare something against what they predict the engine would play.

The most critical part of this is the book, if another skilled centaur prepares a book that allows the engine to reach a position where the drawing moves can be found at the time control, then the engine would be able to hold games to a draw (perhaps all of them). Then again, if the first centaur is able to take a peek at the book, they'd be able to find a hole and exploit it to win a game, so the string of moves that beat those engine exist.

And from what has been shown, maybe you don't need a strong centaur at the other side, as Leela nets keep progressing Stockfish's loses keep increasing. Back when Alpha Zero appeared some people thought it was close to perfection, so they'd have expected it'd have been able to play perfect chess given some big hardware or time control, but we know long time ago that strength has been matched, and surpassed, and it was nothing special.

In 2025 most computer chess enthusiast will have common computers that will be faster than today's $15k Workstations, and software advancements will have continued to grow exponentially. But the horizon effect will still have games being decided, and the strength of today's programs on those only look impressive because we haven't seen something that strong before, in the future we will see them as nothing special, just like clusters with top engines of the past would get beaten badly by my 10 years old computer with current software.

[CMCanavessi]

Unassisted engines are still far away from top correspondence players

Are unassisted engines far away from top "unassisted" correspondence players? I don't think so. I think an engine like SF would win 100% of the games against the best correspondence player in the world, if he can't use engines for analysis.

[Spliffjiffer]

definately true from my pov...the only point of truth a corr-player is able to get is a combination of the "not questionable output" that the engine is able to output together with a human intuition and investigation time he is able to invest...truth, even in chess, is sth that relys on a believe when it comes to the best centaurs of our days...this will change when everything will be bean-counted oc...this will be a TB !...
but even a TB will not secure an approximate way of perfect chess...u will need the most complex way to make your opp blunder somewhere, i believe !?

[Ovyron]

Are unassisted engines far away from top "unassisted" correspondence players?

Top unassisted corr players max out at some 2200 ICCF elo, I believe top engine slaves (who would just put the engine to analyze and spit out their moves without interaction) top at 2400 elo, though it may be because they play too many games (if they only played a few games their rating would be higher.)

So unassisted engines are already ahead of unassisted corr players, but I was talking about assisted corr players (in games where they know they'e playing against unassisted engines and can lead to positions that unassisted engines misplay.)

[hgm]

I don't think a pure 32-men EGT would achieve better than 50%. It would recognize that the initial position is a draw, and would start randomly playing moves that preserve that draw. Which almost always will worsen its position. This will continue until it is at the brink of being lost. Then it will be forced to play the best move, as it will be the only non-losing move it has, so that it will manage to secure the draw. It will never get anywhere near an advantage that is so high that current top engines could make a mistake so large that they give away the game.

To put a fallible opponent under pressure, so that he might make a losing mistake, requires being able to distinguish draws where you were better from draws where you are worse. Such info is not in the EGT; you will need an engine with a heuristic evaluation for that.

[Alayan]

HGM just summed it up perfectly.

[mmt]

I don't think a pure 32-men EGT would achieve better than 50%. It would recognize that the initial position is a draw, and would start randomly playing moves that preserve that draw. Which almost always will worsen its position. This will continue until it is at the brink of being lost. Then it will be forced to play the best move, as it will be the only non-losing move it has, so that it will manage to secure the draw. It will never get anywhere near an advantage that is so high that current top engines could make a mistake so large that they give away the game.

To put a fallible opponent under pressure, so that he might make a losing mistake, requires being able to distinguish draws where you were better from draws where you are worse. Such info is not in the EGT; you will need an engine with a heuristic evaluation for that.

If we go with just the current type of EGTBs it makes sense. It couldn't make too many pointless moves with black though, as white could get a winning advantage and that would put the engines under quite a bit of pressure. It could move to a better score for black according to engines just as it could move to a lower score, however. I believe there are still white losses among top engines playing each other though, so we know that exactly 50% is incorrect. My guess is it would be significantly higher than 50%. We could do a test like I suggested and approximately find out.

[hgm]

It would be above 50%, but I would be surprised if it was above 50.5%. The number of poor moves is usually a lot larger than the number of good moves, so if it not close to the brink of losing (so that most poor moves will be rejected), it will deteriorate its position pretty fast. The draw margin in Chess is so wide that the probability that it will do enough moves that improve its position in a row to get to a won position should be very, very small.