M ANSARI wrote: ↑Thu Mar 07, 2019 9:24 am
I think if you wanted to create a network for Lc0 that plays really well at Knight odds is to train it to play without the Knight. I think it will very quickly figure out the best chances in the position and play much stronger than if it was simply playing with a network trained on the normal start position. This is why I also think that Lc0 will soon be unbeatable at Chess 960. If it took Google 4 hours to get a strong network at the start position ... theoretically once hardware catches up you could have a network for each of the Chess 960 positions and if the few weaknesses of Lc0 are sorted out, this would make a formidable Chess 960 engine as it would play each of the start positions equal to how it now plays the classic chess start position.
stockfish of today will not be able to beat lc0.
The stockfish framework does not go in the right way to beat lc0.
It does not mean that lc0 is going to be unbeatable.
M ANSARI wrote: ↑Thu Mar 07, 2019 9:24 am
I think if you wanted to create a network for Lc0 that plays really well at Knight odds is to train it to play without the Knight. I think it will very quickly figure out the best chances in the position and play much stronger than if it was simply playing with a network trained on the normal start position. This is why I also think that Lc0 will soon be unbeatable at Chess 960. If it took Google 4 hours to get a strong network at the start position ... theoretically once hardware catches up you could have a network for each of the Chess 960 positions and if the few weaknesses of Lc0 are sorted out, this would make a formidable Chess 960 engine as it would play each of the start positions equal to how it now plays the classic chess start position.
I don't think it is necessary to train Lc0 at knight odds for it to play well, it seems that at least network 11248 is already incredibly good at it! I've played a bunch of 5' + 5" games against it, alternating between b1 and g1 knight odds, and setting TempValueCutoff for 0.2 for 3 moves to avoid repeat openings (at little cost since it thinks it has about 12% win chance at the odds, so dropping to 11.8% occasionally on the first three moves is minor), and I've lost every game. No trivial blunders or time losses. I normally beat A/B engines under these conditions. I am truly amazed. Also this is only using a 1060 GPU; I should have a 2080 pretty soon but I can't do any worse! There are limits though; I won quite easily with the same conditions at rook odds. I think that training the network at knight odds wouldn't help very much, as the odds giver would lose almost every game. Maybe if the games were played with a large time (or nodes) handicap to produce fairly even results this would indeed work well.
I encourage any readers above 2000 elo strength to try the same test if they have a 1060 or better GPU. My own strength in fast games is probably only around 2200 FIDE level now, so if you are much stronger than that at blitz perhaps you will fare better than I did. But be sure to use 11248 (or one close to it); recent networks are not nearly as good at giving knight odds, some are awful.
I think that it may be interesting what is the maximal odd that you can give to engine and still draw or win.
It is going to be good to have a match between engines when one side(starting with all the pieces) is deterministic komodo or stockfish or lc0 at fixed level so it is forced to play exactly the same moves in every game.
The second side is the odd giver and should be programmed based on the knowledge that the opponent has to repeat the same moves so it can use it for pruning(for example if you know that the opponent is going to reply 1.d4 by 1.d5 then you can prune in the search other replies of black in the games that you play).
Possible type of questions:
1)How many nodes per move engines need to be able never to lose with white when the opponent start without knight b8?
2)How many nodes per move engines need to be able always to win with black white play without rook h1?
It may be interesting to have a tool that help to answer this question(and at least give final no for different cases)
We do not know final yes because the fact that the engine always won when the opponent is without rook h1 does not prove that it is impossible to draw or win against it.
I got some solid data bearing on the question. Two days ago I got a new laptop with a 2080 GPU. Although the networks starting with 3,4, or 5 are hopeless when giving knight odds, Lc0 network 11248 has no problem. I have played six games at 5' + 5" against it at knight odds, alternating b1 and g1 knights and varying White's opening move among the best options. I lost every game, most were not even close. Then my son Ray played four games against it with the same odds, two at 5' + 5" and two and 10' + 10". Again Lc0 won every game, and on the board, not on time! Ray is an IM, rated higher than I am now, and tied for the Canadian Open championship last year, so he's not a bad player. Now I'm sure we would do better with say 45' + 15" time limit, and of course a new GM with the 2500 minimum FIDE rating would also do better, but given the zero for ten result I would imagine that the 2500 player at the 45' + 15" time limit would at least have a fairly competitive match at knight odds. Probably a top GM wouldn't give up more than an occasional draw at that level, but let's see how things stand a year or so from now.
I think that it is interesting to see games and maybe humans can learn from it better how to play better with material disadvantage against other humans.
I thought in the past that in the high level you can resign if you lose a piece and I remember that I tried some years ago to play with knight odd against a human in correspondence chess with rybka when I trusted the human not to use a computer and he won convincingly and the human was only rating near 2000 and I believe him that he did not use more than an hour for all the game inspite of the fact that it was a correspondence game.
Here is the game from 2007.
I changed move 1 to play different and played 1.f4 for variety relative to previous game against a different opponent and also changed move 10 because I did not want to allow trading pieces but maybe it was a mistake.
other moves are rybka's move because I had no idea how to change rybka's move
I decided to resign at move 39.
[pgn]
[Event "Friend mode"]
[Site "?"]
[Date "2007.10.12"]
[Round "?"]
[White "-"]
[Black "-"]
[Result "*"]
[BlackElo "2400"]
[WhiteElo "2400"]
[TimeControl "1/259200:0"]
[SetUp "1"]
[FEN "rnbqkbnr/pppppppp/8/8/8/8/PPPPPPPP/R1BQKBNR w KQkq - 0 1"]
[Termination "unterminated"]
[PlyCount "76"]
[WhiteType "human"]
[BlackType "human"]
{-------------- r n b q k b n r p p p p p p p p . . . . . . . . . . . . . .
. . . . . . . . . . . . . . . . . . P P P P P P P P R . B Q K B N R white
to play -------------- -------------- r n b q k b n r p p p p p p p p . . .
. . . . . . . . . . . . . . . . . . . . . . . . . . . . . P P P P P P P P R
. B Q K B N R white to play -------------- 512MB, RybkaII.ctg, URI-AMD} 1.
f4 Nf6 2. Nf3 d5 3. e3 c5 4. b3 e6 5. Be2 Bd6 6. O-O Nc6 7. Bb2 Qe7 8. Ne5
O-O 9. Bb5 Qc7 10. Nf3 Be7 11. Bxc6 Qxc6 12. Ne5 Qe8 13. d3 Nd7 14. Nf3 Bd6
15. Qd2 b6 16. e4 Bb7 17. Rae1 d4 18. e5 Be7 19. Bc1 f5 20. exf6 Bxf6 21.
Qe2 Bd5 22. Nd2 Qg6 23. Ne4 Bh4 24. g3 Be7 25. Bd2 Rae8 26. Qf3 Nf6 27. Qe2
Nxe4 28. dxe4 Bb7 29. Rf2 Rd8 30. Qd3 Qe8 31. a3?? Qc6 32. h4 b5 33. Ba5 c4
34. bxc4 bxc4 35. Qd1 Rd7 36. Bb4 Bxb4 37. axb4 c3 38. Rf3 Rfd8 *[/pgn]
lkaufman wrote: ↑Mon Mar 18, 2019 11:51 pm
but given the zero for ten result I would imagine that the 2500 player at the 45' + 15" time limit would at least have a fairly competitive match at knight odds. Probably a top GM wouldn't give up more than an occasional draw at that level, but let's see how things stand a year or so from now.
Yes, I also got by some sort of simulation that a top GM in these conditions would win 7-8 games out of 10, but will draw (or even lose) 2-3 games. I think Komodo with a specific Contempt or even specific eval for Knight odds, can perform similarly. Some matches can be organized against top GMs:
Human (top GM) gets (Wins - Draws or Losses) reward in say hundreds of $$$, so it would be important for a human to Win as many as possible at Knight odds. The goal of this thread was to show that Knight odds are already possible against a top GM, with some excitement as the games and the result go.
lkaufman wrote: ↑Mon Mar 18, 2019 11:51 pm
but given the zero for ten result I would imagine that the 2500 player at the 45' + 15" time limit would at least have a fairly competitive match at knight odds. Probably a top GM wouldn't give up more than an occasional draw at that level, but let's see how things stand a year or so from now.
Yes, I also got by some sort of simulation that a top GM in these conditions would win 7-8 games out of 10, but will draw (or even lose) 2-3 games. I think Komodo with a specific Contempt or even specific eval for Knight odds, can perform similarly. Some matches can be organized against top GMs:
Human (top GM) gets (Wins - Draws or Losses) reward in say hundreds of $$$, so it would be important for a human to Win as many as possible at Knight odds. The goal of this thread was to show that Knight odds are already possible against a top GM, with some excitement as the games and the result go.
I would like to see some computer rating lists test engines without a knight to see what rating they can get.
It may be interesting if some version of lc0 can get higher rating relative to a-b engines.
lkaufman wrote: ↑Mon Mar 18, 2019 11:51 pm
but given the zero for ten result I would imagine that the 2500 player at the 45' + 15" time limit would at least have a fairly competitive match at knight odds. Probably a top GM wouldn't give up more than an occasional draw at that level, but let's see how things stand a year or so from now.
Yes, I also got by some sort of simulation that a top GM in these conditions would win 7-8 games out of 10, but will draw (or even lose) 2-3 games. I think Komodo with a specific Contempt or even specific eval for Knight odds, can perform similarly. Some matches can be organized against top GMs:
Human (top GM) gets (Wins - Draws or Losses) reward in say hundreds of $$$, so it would be important for a human to Win as many as possible at Knight odds. The goal of this thread was to show that Knight odds are already possible against a top GM, with some excitement as the games and the result go.
I would like to see some computer rating lists test engines without a knight to see what rating they can get.
It may be interesting if some version of lc0 can get higher rating relative to a-b engines.
I ran a ten game knight odds match at "slow bullet" (1' + 1") between Lc0 (11248) and Rybka 2.3.2a 32 bit (one thread). Lc0 won by 6 to 4! Alternating b1,g1 knights, five best White opening moves (my judgment consulting engines) for each knight. For those who don't know or remember, Rybka 2.3.2a was the top engine in 2007, was the last of the Rybka 2 series, was partly my own work (the evaluation), defeated a former top-5 GM giving him pawn odds by 5.5 to 2.5, and even on one thread and 32 bit is rated 2922 by CCRL blitz. There is little doubt that it would defeat Magnus Carlsen in a standard match, and would win nearly every game from him at 1' + 1", so this result is really remarkable. Lc0 plays knight odds completely differently than A/B engines; it attacks like a madman, throwing pawns forward right from the start with hardly any concern for development, creating chaotic positions where it's hard to just play routine moves. If I take knight odds from Stockfish (with max contempt set) the only issue is whether it can reach a position where it is hard for me to make progress; with Lc0 it is playing to win from the first move! It is unfortunate that newer, stronger Lc0 networks just fall apart at knight odds; even with an initial eval in the four to five percent range they can't play sensibly. I think that this problem is just the other side of the "trolling" problem where Lc0 looks for the longest win rather than the shortest one.
lkaufman wrote: ↑Mon Mar 18, 2019 11:51 pm
but given the zero for ten result I would imagine that the 2500 player at the 45' + 15" time limit would at least have a fairly competitive match at knight odds. Probably a top GM wouldn't give up more than an occasional draw at that level, but let's see how things stand a year or so from now.
Yes, I also got by some sort of simulation that a top GM in these conditions would win 7-8 games out of 10, but will draw (or even lose) 2-3 games. I think Komodo with a specific Contempt or even specific eval for Knight odds, can perform similarly. Some matches can be organized against top GMs:
Human (top GM) gets (Wins - Draws or Losses) reward in say hundreds of $$$, so it would be important for a human to Win as many as possible at Knight odds. The goal of this thread was to show that Knight odds are already possible against a top GM, with some excitement as the games and the result go.
I would like to see some computer rating lists test engines without a knight to see what rating they can get.
It may be interesting if some version of lc0 can get higher rating relative to a-b engines.
I ran a ten game knight odds match at "slow bullet" (1' + 1") between Lc0 (11248) and Rybka 2.3.2a 32 bit (one thread). Lc0 won by 6 to 4! Alternating b1,g1 knights, five best White opening moves (my judgment consulting engines) for each knight. For those who don't know or remember, Rybka 2.3.2a was the top engine in 2007, was the last of the Rybka 2 series, was partly my own work (the evaluation), defeated a former top-5 GM giving him pawn odds by 5.5 to 2.5, and even on one thread and 32 bit is rated 2922 by CCRL blitz. There is little doubt that it would defeat Magnus Carlsen in a standard match, and would win nearly every game from him at 1' + 1", so this result is really remarkable. Lc0 plays knight odds completely differently than A/B engines; it attacks like a madman, throwing pawns forward right from the start with hardly any concern for development, creating chaotic positions where it's hard to just play routine moves. If I take knight odds from Stockfish (with max contempt set) the only issue is whether it can reach a position where it is hard for me to make progress; with Lc0 it is playing to win from the first move! It is unfortunate that newer, stronger Lc0 networks just fall apart at knight odds; even with an initial eval in the four to five percent range they can't play sensibly. I think that this problem is just the other side of the "trolling" problem where Lc0 looks for the longest win rather than the shortest one.
I can confirm the result against Rybka 2.3.2a 32bit (one thread) wit that against Arasan 20.0 64bit (1 thread), about 60 Elo points better rated engine (almost 3000 CCRL 40/4 Elo points). My GPU is RTX 2070, a bit weaker than your RTX 2080. At 60s+1s time control, Lc0 ID11248 got 2 wins and 2 draws in 10 games, for 3.0/10 result at Knight odds. Which is in the same ballpark with your result, accounting for stronger Arasan and weaker GPU. For variety, I just set temperature to 1.5 for 3 starting moves of Lc0.
