Larry Kaufman

[lkaufman]

Dear GM Larry Kaufman,

Please accept my belated birthday wishes for your continued chess, chess engine, and personal successes.

All the very best!

ICCF GM - Stephen Ham

Thanks. At an age where most people (other than recent U.S. Presidents and Senators!) have retired, things are going unusually well for me, with Komodo Dragon improving at its fastest clip in a decade or more. I could never have imagined it reaching the current level, and even less imagined that the rate of improvement would be accelerating!

[dangi12012]

Thanks. At an age where most people (other than recent U.S. Presidents and Senators!) have retired, things are going unusually well for me, with Komodo Dragon improving at its fastest clip in a decade or more. I could never have imagined it reaching the current level, and even less imagined that the rate of improvement would be accelerating!

So mr. kaufman - I have 2 questions:
1) Do you see any advantages in calculating the 8 man tablebase which is technically possible in 2022 but a huge time investment programming/cpu time?

2) Wouldnt it be fun to have an engine where you could change what it thinks a win is? For example stockfish that thinks getting mated in 6 is a win. Then it would always set itself up for defeat in 6 moves wherever possible. It would be a cool training opportunity where mate in 6 puzzles would constantly emerge during natural play.
Or Stockfish that thinks draw is a win. It would be very forcefully find drawing opportunities and never go for real checkmate. Could be fun too!

[lkaufman]

Thanks. At an age where most people (other than recent U.S. Presidents and Senators!) have retired, things are going unusually well for me, with Komodo Dragon improving at its fastest clip in a decade or more. I could never have imagined it reaching the current level, and even less imagined that the rate of improvement would be accelerating!

So mr. kaufman - I have 2 questions:
1) Do you see any advantages in calculating the 8 man tablebase which is technically possible in 2022 but a huge time investment programming/cpu time?

2) Wouldnt it be fun to have an engine where you could change what it thinks a win is? For example stockfish that thinks getting mated in 6 is a win. Then it would always set itself up for defeat in 6 moves wherever possible. It would be a cool training opportunity where mate in 6 puzzles would constantly emerge during natural play.
Or Stockfish that thinks draw is a win. It would be very forcefully find drawing opportunities and never go for real checkmate. Could be fun too!

8 man tablebases would have some added value, but compared with other developments in computer chess it is insignificant and only interesting for answering questions such as whether an extra knight or bishop is generally enough to win with 3 other identical pieces (no pawns) each per side. Regarding your second question, Komodo (and Dragon) already has a UCI option to define draws as a win for either side. We have also made experimental versions that define Stalemate as a win, bare king as a loss, and repetition as a loss. We have talked about making a version, perhaps for a chess.com show, which would require Komodo to mate with a specified piece or piece type, or on a specified square, or both. Many possibilities.

[dangi12012]

Thanks. At an age where most people (other than recent U.S. Presidents and Senators!) have retired, things are going unusually well for me, with Komodo Dragon improving at its fastest clip in a decade or more. I could never have imagined it reaching the current level, and even less imagined that the rate of improvement would be accelerating!

So mr. kaufman - I have 2 questions:
1) Do you see any advantages in calculating the 8 man tablebase which is technically possible in 2022 but a huge time investment programming/cpu time?

2) Wouldnt it be fun to have an engine where you could change what it thinks a win is? For example stockfish that thinks getting mated in 6 is a win. Then it would always set itself up for defeat in 6 moves wherever possible. It would be a cool training opportunity where mate in 6 puzzles would constantly emerge during natural play.
Or Stockfish that thinks draw is a win. It would be very forcefully find drawing opportunities and never go for real checkmate. Could be fun too!

8 man tablebases would have some added value, but compared with other developments in computer chess it is insignificant and only interesting for answering questions such as whether an extra knight or bishop is generally enough to win with 3 other identical pieces (no pawns) each per side. Regarding your second question, Komodo (and Dragon) already has a UCI option to define draws as a win for either side. We have also made experimental versions that define Stalemate as a win, bare king as a loss, and repetition as a loss. We have talked about making a version, perhaps for a chess.com show, which would require Komodo to mate with a specified piece or piece type, or on a specified square, or both. Many possibilities.

My gut feeling tells me that the elo gain by n-Man tablebase will be hyperexponential - since the 32 man tablebase will be chess god - and 0 to 7 brings only 20 Elo points. Whats interesting is that its totally possible today to generate 8 man tablebases with some time investment. Are you interested in that endeavour? I have the funding for the hardware - since its exactly known that you need 110x everything thats needed in 7 man TB.
But with CUDA and some fast NVME SSDS you dont even need so much RAM anymore.

[lkaufman]

Thanks. At an age where most people (other than recent U.S. Presidents and Senators!) have retired, things are going unusually well for me, with Komodo Dragon improving at its fastest clip in a decade or more. I could never have imagined it reaching the current level, and even less imagined that the rate of improvement would be accelerating!

So mr. kaufman - I have 2 questions:
1) Do you see any advantages in calculating the 8 man tablebase which is technically possible in 2022 but a huge time investment programming/cpu time?

2) Wouldnt it be fun to have an engine where you could change what it thinks a win is? For example stockfish that thinks getting mated in 6 is a win. Then it would always set itself up for defeat in 6 moves wherever possible. It would be a cool training opportunity where mate in 6 puzzles would constantly emerge during natural play.
Or Stockfish that thinks draw is a win. It would be very forcefully find drawing opportunities and never go for real checkmate. Could be fun too!

8 man tablebases would have some added value, but compared with other developments in computer chess it is insignificant and only interesting for answering questions such as whether an extra knight or bishop is generally enough to win with 3 other identical pieces (no pawns) each per side. Regarding your second question, Komodo (and Dragon) already has a UCI option to define draws as a win for either side. We have also made experimental versions that define Stalemate as a win, bare king as a loss, and repetition as a loss. We have talked about making a version, perhaps for a chess.com show, which would require Komodo to mate with a specified piece or piece type, or on a specified square, or both. Many possibilities.

My gut feeling tells me that the elo gain by n-Man tablebase will be hyperexponential - since the 32 man tablebase will be chess god - and 0 to 7 brings only 20 Elo points. Whats interesting is that its totally possible today to generate 8 man tablebases with some time investment. Are you interested in that endeavour? I have the funding for the hardware - since its exactly known that you need 110x everything thats needed in 7 man TB.
But with CUDA and some fast NVME SSDS you dont even need so much RAM anymore.

I am interested in the endeavor, but I don't agree about the hyperexponential gains. I suspect that the percentage of 8 man positions where the best engines get it wrong is pretty small. It's also not clear what elo remains even with a 32 man TB. If you define Elo based on using opening books of openings currently used by top GMs or correspondence GMs, and assume decent hardware and Rapid (not blitz) time controls, a 32 man tb might not even add a hundred elo. With unbalanced human openings near the win/draw line, then it could be worth a thousand elo or more. The only practical benefit of 8 man TB would be for tournaments like CCC or TCEC where the hardware might support it, that would be up to the organizers of those events. But it would be quite interesting for me personally to be able to answer certain questions, in particular with 8 man the question of how often and under what circumstances does an extra Exchange suffice to win?

[dangi12012]

I am interested in the endeavor, but I don't agree about the hyperexponential gains. I suspect that the percentage of 8 man positions where the best engines get it wrong is pretty small. It's also not clear what elo remains even with a 32 man TB. If you define Elo based on using opening books of openings currently used by top GMs or correspondence GMs, and assume decent hardware and Rapid (not blitz) time controls, a 32 man tb might not even add a hundred elo. With unbalanced human openings near the win/draw line, then it could be worth a thousand elo or more. The only practical benefit of 8 man TB would be for tournaments like CCC or TCEC where the hardware might support it, that would be up to the organizers of those events. But it would be quite interesting for me personally to be able to answer certain questions, in particular with 8 man the question of how often and under what circumstances does an extra Exchange suffice to win?

Im sure that when you setup an 8 man position the time SF cannot solve it is almost zero. The point is that these positions come up during search from 20 plies away - and then you trade 1 lookup for billions of nodes that need not be searched. Effectively chess ends for the engine with 8 pieces left on the board - and not with actual mate.

Im suprised that 6 man adds only 20 elo - but i found no current data for 7 man (which should exist somewhere).
My point is that when N = 6 => 20 elo and N = 32 => 4000+ ELO the relationship cannot be linear.

I dont know what you mean with "32 man tb migh not add elo". Since it would be a perfect chessplayer so at least 4000 Elo or however much that will end up being. It looks small but its actually a logarithmic relationship - so +100elo is an insane gain in terms of win/lose ratio.

[lkaufman]

I am interested in the endeavor, but I don't agree about the hyperexponential gains. I suspect that the percentage of 8 man positions where the best engines get it wrong is pretty small. It's also not clear what elo remains even with a 32 man TB. If you define Elo based on using opening books of openings currently used by top GMs or correspondence GMs, and assume decent hardware and Rapid (not blitz) time controls, a 32 man tb might not even add a hundred elo. With unbalanced human openings near the win/draw line, then it could be worth a thousand elo or more. The only practical benefit of 8 man TB would be for tournaments like CCC or TCEC where the hardware might support it, that would be up to the organizers of those events. But it would be quite interesting for me personally to be able to answer certain questions, in particular with 8 man the question of how often and under what circumstances does an extra Exchange suffice to win?

Im sure that when you setup an 8 man position the time SF cannot solve it is almost zero. The point is that these positions come up during search from 20 plies away - and then you trade 1 lookup for billions of nodes that need not be searched. Effectively chess ends for the engine with 8 pieces left on the board - and not with actual mate.

Im suprised that 6 man adds only 20 elo - but i found no current data for 7 man (which should exist somewhere).
My point is that when N = 6 => 20 elo and N = 32 => 4000+ ELO the relationship cannot be linear.

I dont know what you mean with "32 man tb migh not add elo". Since it would be a perfect chessplayer so at least 4000 Elo or however much that will end up being. It looks small but its actually a logarithmic relationship - so +100elo is an insane gain in terms of win/lose ratio.

My point is that "adds 20 elo" (or any other number) has no meaning anymore unless you define the conditions of the test (opening book, time limit, threads). The same change can come out as 10 or 100 elo depending on how the test is done. Maybe this is outside the topic of 8 man tb, but it is unavoidable if you quote elo differences. If you mean that the win to loss ratio gains more and more with each added man in N man tb, you may be correct, but the actual elo gain probably declines due to more and more draws. Bottom line: 8 man tb (if you can access it quickly) will help results just like a small speedup in the engine will help results, but whether this is meaningful in terms of Elo depends on the test conditions.

[dangi12012]

I am interested in the endeavor, but I don't agree about the hyperexponential gains. I suspect that the percentage of 8 man positions where the best engines get it wrong is pretty small. It's also not clear what elo remains even with a 32 man TB. If you define Elo based on using opening books of openings currently used by top GMs or correspondence GMs, and assume decent hardware and Rapid (not blitz) time controls, a 32 man tb might not even add a hundred elo. With unbalanced human openings near the win/draw line, then it could be worth a thousand elo or more. The only practical benefit of 8 man TB would be for tournaments like CCC or TCEC where the hardware might support it, that would be up to the organizers of those events. But it would be quite interesting for me personally to be able to answer certain questions, in particular with 8 man the question of how often and under what circumstances does an extra Exchange suffice to win?

Im sure that when you setup an 8 man position the time SF cannot solve it is almost zero. The point is that these positions come up during search from 20 plies away - and then you trade 1 lookup for billions of nodes that need not be searched. Effectively chess ends for the engine with 8 pieces left on the board - and not with actual mate.

Im suprised that 6 man adds only 20 elo - but i found no current data for 7 man (which should exist somewhere).
My point is that when N = 6 => 20 elo and N = 32 => 4000+ ELO the relationship cannot be linear.

I dont know what you mean with "32 man tb migh not add elo". Since it would be a perfect chessplayer so at least 4000 Elo or however much that will end up being. It looks small but its actually a logarithmic relationship - so +100elo is an insane gain in terms of win/lose ratio.

My point is that "adds 20 elo" (or any other number) has no meaning anymore unless you define the conditions of the test (opening book, time limit, threads). The same change can come out as 10 or 100 elo depending on how the test is done. Maybe this is outside the topic of 8 man tb, but it is unavoidable if you quote elo differences. If you mean that the win to loss ratio gains more and more with each added man in N man tb, you may be correct, but the actual elo gain probably declines due to more and more draws. Bottom line: 8 man tb (if you can access it quickly) will help results just like a small speedup in the engine will help results, but whether this is meaningful in terms of Elo depends on the test conditions.

I quote the internet:
Elo is not a measurement like height or speed. It's only useful in certain conditions, like opponents are not hugely mismatched, the players in the pool play often, and a variety of opponents. When this happens the rating is a good predictor for your performance within that population.
‐------------------------------------------
That being said a reference 3000 elo engine would be timeless and deterministic. Then you could play your own against it (time control only for self) and get the reference elo score.

As i see it now only some chessbots online have that that or am i mistaken?

[lkaufman]

I am interested in the endeavor, but I don't agree about the hyperexponential gains. I suspect that the percentage of 8 man positions where the best engines get it wrong is pretty small. It's also not clear what elo remains even with a 32 man TB. If you define Elo based on using opening books of openings currently used by top GMs or correspondence GMs, and assume decent hardware and Rapid (not blitz) time controls, a 32 man tb might not even add a hundred elo. With unbalanced human openings near the win/draw line, then it could be worth a thousand elo or more. The only practical benefit of 8 man TB would be for tournaments like CCC or TCEC where the hardware might support it, that would be up to the organizers of those events. But it would be quite interesting for me personally to be able to answer certain questions, in particular with 8 man the question of how often and under what circumstances does an extra Exchange suffice to win?

Im sure that when you setup an 8 man position the time SF cannot solve it is almost zero. The point is that these positions come up during search from 20 plies away - and then you trade 1 lookup for billions of nodes that need not be searched. Effectively chess ends for the engine with 8 pieces left on the board - and not with actual mate.

Im suprised that 6 man adds only 20 elo - but i found no current data for 7 man (which should exist somewhere).
My point is that when N = 6 => 20 elo and N = 32 => 4000+ ELO the relationship cannot be linear.

I dont know what you mean with "32 man tb migh not add elo". Since it would be a perfect chessplayer so at least 4000 Elo or however much that will end up being. It looks small but its actually a logarithmic relationship - so +100elo is an insane gain in terms of win/lose ratio.

My point is that "adds 20 elo" (or any other number) has no meaning anymore unless you define the conditions of the test (opening book, time limit, threads). The same change can come out as 10 or 100 elo depending on how the test is done. Maybe this is outside the topic of 8 man tb, but it is unavoidable if you quote elo differences. If you mean that the win to loss ratio gains more and more with each added man in N man tb, you may be correct, but the actual elo gain probably declines due to more and more draws. Bottom line: 8 man tb (if you can access it quickly) will help results just like a small speedup in the engine will help results, but whether this is meaningful in terms of Elo depends on the test conditions.

I quote the internet:
Elo is not a measurement like height or speed. It's only useful in certain conditions, like opponents are not hugely mismatched, the players in the pool play often, and a variety of opponents. When this happens the rating is a good predictor for your performance within that population.
‐------------------------------------------
That being said a reference 3000 elo engine would be timeless and deterministic. Then you could play your own against it (time control only for self) and get the reference elo score.

As i see it now only some chessbots online have that that or am i mistaken?

I don't understand the question. Every rating list has one (or more) reference engines with a defined elo, but this is not interesting to talk about here, we are only talking about elo GAINS, so the reference elo doesn't matter. The point is that anything you read about elo that is more than a year or so old is pretty much obsolete (with respect to top engines). The current situation is that the top engines on many cores with non-blitz time controls draw almost all their games with each other unless forced to play bad openings. So the elo gain from any change (to engine, to nets, or to TBs) is now hugely depending on the opening book used to get the needed variety. This wasn't a big issue a few years ago. Most engine rating lists (other than one-person lists) leave the book choice up to the tester, so the standard is undefined.

[Leo]

Thanks. At an age where most people (other than recent U.S. Presidents and Senators!) have retired, things are going unusually well for me, with Komodo Dragon improving at its fastest clip in a decade or more. I could never have imagined it reaching the current level, and even less imagined that the rate of improvement would be accelerating!

So mr. kaufman - I have 2 questions:
1) Do you see any advantages in calculating the 8 man tablebase which is technically possible in 2022 but a huge time investment programming/cpu time?

2) Wouldnt it be fun to have an engine where you could change what it thinks a win is? For example stockfish that thinks getting mated in 6 is a win. Then it would always set itself up for defeat in 6 moves wherever possible. It would be a cool training opportunity where mate in 6 puzzles would constantly emerge during natural play.
Or Stockfish that thinks draw is a win. It would be very forcefully find drawing opportunities and never go for real checkmate. Could be fun too!

8 man tablebases would have some added value, but compared with other developments in computer chess it is insignificant and only interesting for answering questions such as whether an extra knight or bishop is generally enough to win with 3 other identical pieces (no pawns) each per side. Regarding your second question, Komodo (and Dragon) already has a UCI option to define draws as a win for either side. We have also made experimental versions that define Stalemate as a win, bare king as a loss, and repetition as a loss. We have talked about making a version, perhaps for a chess.com show, which would require Komodo to mate with a specified piece or piece type, or on a specified square, or both. Many possibilities.

My gut feeling tells me that the elo gain by n-Man tablebase will be hyperexponential - since the 32 man tablebase will be chess god - and 0 to 7 brings only 20 Elo points. Whats interesting is that its totally possible today to generate 8 man tablebases with some time investment. Are you interested in that endeavour? I have the funding for the hardware - since its exactly known that you need 110x everything thats needed in 7 man TB.
But with CUDA and some fast NVME SSDS you dont even need so much RAM anymore.

I found this information a few years ago. Quote:
EGTB
6 man tablebase syzygy {EGTB} 90 GB
7 man tablebase The high speed of generating the tablebases was because of using a supercomputer named Lomonosov (top500). The size of seven-man tablebases is about 140 TB.

Many show interest in what is to expect from 8-man endings. First, take note that the longest 6-man mate took 262 moves (KRN-KNN). Moving to 7-man endings doubled this value. Second, 8-man tablebases include much more endings with both sides having relatively equal strength. All this gives us a strong hope to discover a mate in more than 1000 moves in one of 8-man endgames. Unfortunately the size of 8-man tablebases will be 100 times larger than the size of 7-man tablebases. To fully compute them, one will need about 10 PB (10,000 TB) of disk space and 50 TB of RAM. Only the top 10 supercomputers can solve the 8-man problem in 2014. The first 1000-move mate is unlikely to be found until 2020 when a part of a TOP100 supercomputer may be allowed to be used for solving this task.