Perfect chess engine elo ( 32 men TB) can be within 200 of Stocfish in Tcec LTC conditions

[Nay Lin Tun]

If there were no opening book, Stockfish will be minimally able to draw another Stockfish or Leela in 95% of games. If Stockfish is playing against 32 men TB, the extreme condition would be 50 draw 50 losses , which is within 200 elo.


On the other hand, current Stockfish with contempt may be slightly higher rated than 32 men TB in Round Robin rating pool. Prettty crazy right? Actually 32 men TB doesn't store statistics of winning chances or knowledge of maximising chances to win which will subsequently lose some elo. For example, in KRB vs KR, 7 men TB will randomly throw away one rook as it knows that all moves are draw. ( KB vs KR is still draw). Carry behaviour of how this 7 men TB play into 32 men TB, we can see TB will be randomly choose the first move 1. a3 or f3 or h3 or e4 (assuming that except 1.g4, all other 19 moves are likely to be draw). Subsequently, 32 men TB will choose next move like. 1. a3 e5 2. a4 ( It still knows as a draw but it throw away practical winning chances)

[Dann Corbit]

Twenty years from now the current hardware and software will grind today's TCEC hardware and latest chess software into powder.
Achieving a draw against a peer is not the same thing as achieving a draw against a perfect chess player.

[duncan]

Twenty years from now the current hardware and software will grind today's TCEC hardware and latest chess software into powder.
Achieving a draw against a peer is not the same thing as achieving a draw against a perfect chess player.

Does your prediction account for that top rated chess engines score a much higher percentage of draws against their peers, than they did 20 years ago?

[OneTrickPony]

Maybe it's 200 points away, maybe 400. We just don't know. The trick is going to be to select openings that are very close to losing - either drawn but difficult to hold or won but difficult to convert. TCEC is going into that direction and I feel it will be more than 200 ELO vs godlike engine from the future with such openings.

The starting position and the main line solid openings I think we are closer than 200 ELO though. It's hard to imagine current SF being outplayed from main line Ragozin, Berlin or Open Spanish positions even with all the time and advancements from the future.

[Nay Lin Tun]

Maybe it's 200 points away, maybe 400. We just don't know. The trick is going to be to select openings that are very close to losing - either drawn but difficult to hold or won but difficult to convert. TCEC is going into that direction and I feel it will be more than 200 ELO vs godlike engine from the future with such openings.

The starting position and the main line solid openings I think we are closer than 200 ELO though. It's hard to imagine current SF being outplayed from main line Ragozin, Berlin or Open Spanish positions even with all the time and advancements from the future.

I understand rating can vary with hardwares, time control and opening book.

Yes, in extreme fast time control with extremely unbalanced opening conditions, the rating gap may be within 400 elo. ( 80 losses, 20 draw). However, the rating gap will be narrower like within 100 elo in ICCF conditions (2 weeks per move , TCEC hardware, no opening books).

[Uri Blass]

If there were no opening book, Stockfish will be minimally able to draw another Stockfish or Leela in 95% of games. If Stockfish is playing against 32 men TB, the extreme condition would be 50 draw 50 losses , which is within 200 elo.


On the other hand, current Stockfish with contempt may be slightly higher rated than 32 men TB in Round Robin rating pool. Prettty crazy right? Actually 32 men TB doesn't store statistics of winning chances or knowledge of maximising chances to win which will subsequently lose some elo. For example, in KRB vs KR, 7 men TB will randomly throw away one rook as it knows that all moves are draw. ( KB vs KR is still draw). Carry behaviour of how this 7 men TB play into 32 men TB, we can see TB will be randomly choose the first move 1. a3 or f3 or h3 or e4 (assuming that except 1.g4, all other 19 moves are likely to be draw). Subsequently, 32 men TB will choose next move like. 1. a3 e5 2. a4 ( It still knows as a draw but it throw away practical winning chances)

1)I see no reason to use a stupid 32 men TB even if we have it if it means playing a random draw move.
I suspect that with a random draw moves even I can draw against the 32 men TB in most of the games.

It is better to simply have a memory of practical moves to play in every TB position.
It can be more than one move to prevent humans to easily draw by learning a single game but also less than all drawing moves so it is not going to have moves that do not help to give good practical chances.

2)I am not sure if it is possible to beat stockfish or Leela with no book in TCEC conditions.
Maybe they play well enough to get at least a draw in every game regardless of the moves of the opponent.

If it is not the case(and my guess is that it is not the case) then it will be interesting to see games that prove it.

[mwyoung]

If there were no opening book, Stockfish will be minimally able to draw another Stockfish or Leela in 95% of games. If Stockfish is playing against 32 men TB, the extreme condition would be 50 draw 50 losses , which is within 200 elo.


On the other hand, current Stockfish with contempt may be slightly higher rated than 32 men TB in Round Robin rating pool. Prettty crazy right? Actually 32 men TB doesn't store statistics of winning chances or knowledge of maximising chances to win which will subsequently lose some elo. For example, in KRB vs KR, 7 men TB will randomly throw away one rook as it knows that all moves are draw. ( KB vs KR is still draw). Carry behaviour of how this 7 men TB play into 32 men TB, we can see TB will be randomly choose the first move 1. a3 or f3 or h3 or e4 (assuming that except 1.g4, all other 19 moves are likely to be draw). Subsequently, 32 men TB will choose next move like. 1. a3 e5 2. a4 ( It still knows as a draw but it throw away practical winning chances)

The 32 man TB does not need to be smart. But there is a trick that has worked in the past. And that would most likely work with a 32 man TB. And that is to use the TB data to play the move that makes the game last the longest in a draw score position for move 1 with a 32 man TB. When playing against the current TCEC hardware. Chess games are won by mistakes of the weaker engine. The more moves that are played in a game. The more chance the current TCEC engines will make an error. I think you would be surprised how good this 32 man TB trick can work against a non perfect player.

[BrendanJNorman]

Talking about "perfect chess" or 32 piece TBs is like talking about aliens or similar.

We have next to no data to analyze any of these things.

Given that Stockfish plays far from perfect chess (very strong is still far from perfect, just as 7 piece tablebases are exponentially far away from 32 piece...imo), my feeling is that it would lose 100-0 to 32 piece tablebases or a "perfect chess" entity.

[duncan]

Talking about "perfect chess" or 32 piece TBs is like talking about aliens or similar.

We have next to no data to analyze any of these things.

Given that Stockfish plays far from perfect chess (very strong is still far from perfect, just as 7 piece tablebases are exponentially far away from 32 piece...imo), my feeling is that it would lose 100-0 to 32 piece tablebases or a "perfect chess" entity.

Do you have an elo estimate for 32 piece. I think Kai would you just give it another 400 elo.

[hgm]

EGT often play moronically stupid in drawn positions. An engine playing only from a 32-men EGTB would of course never lose, but it would also virtually never win. In all its games it would very quickly built up a nearly-losing disadvatage. And only then would it start to fight for a draw, which it would reach through very clever moves. Such an engine would probably not even be at the top of the current CCRL list. It would never lose, but the fact that it would also never win against 2500-rated opponents would severely depress its rating.

EGT are not the answer to everything. You need some form of opponent modeling to select good move in drawn positions. Merely trying to make the game last long in the hope the opponent will err won't do it. It would just lead to pointless shuffling of its own pieces, trying to minimize Pawn moves to keep them for resetting the ply counter each time it approaches 100. Just like minimizing the number of non-losing replies also is silly: that would just lead to quickly trading everything, because it can be recaptured in only few ways, and if the opponent forgets to recapture, you have won. But of course there is zero chance that he will forget to recapture. You somehow have to be able to seek out positions where the opponent is likely to make an error.