Jouni wrote: ↑Tue Nov 17, 2020 3:28 pm
My personal feeling: human computer era ended around 2005 when SMP Rybka was released. Handicap matches are not interesting.
I would like to see non handicap games against GM's even if the engines win every game.
Note that I am not sure that they can win every game with black and maybe humans can find some strategy that give them practical chances for a draw.
I agree. To me, the only thing a handicap match accomplishes is embarrassment to the man if he still loses.
GM Nakamura does not care about such things. He likes the competition and more views to his channel.
Here is GM Nakamura playing Komodo in a 10 game match straight up. And gets crushed 10-0.
It is interesting to note in the video in the last game. GM Nakamura says he thinks a 2 pawn advantage for him. Would be a even match setup.
Yes, and many draws. I am getting in simulations Dragon vs LC0 J92-330 "the worst case scenario" for Dragon as 2 losses and 6 draws. And the best case as 3 wins and 5 draws.
My bets would be:
1/ There will be more than 3 draws
2/ Naka won't get more than 2 wins
Jouni wrote: ↑Tue Nov 17, 2020 3:28 pm
My personal feeling: human computer era ended around 2005 when SMP Rybka was released. Handicap matches are not interesting.
I would like to see non handicap games against GM's even if the engines win every game.
Note that I am not sure that they can win every game with black and maybe humans can find some strategy that give them practical chances for a draw.
I agree. To me, the only thing a handicap match accomplishes is embarrassment to the man if he still loses.
GM Nakamura does not care about such things. He likes the competition and more views to his channel.
Here is GM Nakamura playing Komodo in a 10 game match straight up. And gets crushed 10-0.
It is interesting to note in the video in the last game. GM Nakamura says he thinks a 2 pawn advantage for him. Would be a even match setup.
Yes, and many draws. I am getting in simulations Dragon vs LC0 J92-330 "the worst case scenario" for Dragon as 2 losses and 6 draws. And the best case as 3 wins and 5 draws.
My bets would be:
1/ There will be more than 3 draws
2/ Naka won't get more than 2 wins
It seems that you expect a close match, as do I. How much time (or how many nodes) did you give Lc0 in the simulation? Did you base it on Hikaru's standard FIDE rating or on his Rapid FIDE rating (which is much higher), or something else? How about the hardware for Komodo, is this just assuming we ran on your hardware? When I last tried this, Lc0 was not a good simulator of a human playing handicap, but that was a while ago, perhaps now it is okay. There's also the question of MCTS vs. normal Dragon; probably MCTS will do worse in simulation, but perhaps better against actual human.
I used vague extrapolations, but here a factor of 2, especially in Dragon speed or TC, is hardly changing anything. I took Naka's rating as 2800 FIDE Rapid, and 300-400k SF12 nodes per move needed to match Naka (2800) based on our older 45min + 30s estimates with Komodo and SF (older). At 15min + 10s this would mean at most 150k SF12 nodes per move and that is about 100-200 nodes of Lc0 (large net J92-330) as strength goes (direct match). A factor of 3-4 with Lc0 nodes per move (human TC) is Elo-wise a much larger factor of about 20-40 (from scaling properties, diminishing returns) for Dragon full speed or nodes per move at 15min + 10s. All in all I have reduced the match to the simulation of 50 nodes per move Lc0 (Drawscore=80) and 100s + 1s full strength 4 core Dragon (Contempt=100). While Dragon can be 2-3 time faster or slower without perturbing much the result, the difference with Lc0 at 30 and 100 nodes per move is not negligible, and that's pretty much my worst case and best case results for Naka. And the results show some at least a half draws (many failed Naka wins) and somewhat on average larger Dragon number of wins, but with serious possibility that Naka might win the match.
That is funny. We each had a calabration method, but came to the same conclusion in terms of Lc0 nodes per move. I just used nodes per second limit option set at 2.3. Resulting in the about the same nodes = Naka per move. And both showing a Naka win.
Jouni wrote: ↑Tue Nov 17, 2020 3:28 pm
My personal feeling: human computer era ended around 2005 when SMP Rybka was released. Handicap matches are not interesting.
I would like to see non handicap games against GM's even if the engines win every game.
Note that I am not sure that they can win every game with black and maybe humans can find some strategy that give them practical chances for a draw.
I agree. To me, the only thing a handicap match accomplishes is embarrassment to the man if he still loses.
GM Nakamura does not care about such things. He likes the competition and more views to his channel.
Here is GM Nakamura playing Komodo in a 10 game match straight up. And gets crushed 10-0.
It is interesting to note in the video in the last game. GM Nakamura says he thinks a 2 pawn advantage for him. Would be a even match setup.
Yes, and many draws. I am getting in simulations Dragon vs LC0 J92-330 "the worst case scenario" for Dragon as 2 losses and 6 draws. And the best case as 3 wins and 5 draws.
My bets would be:
1/ There will be more than 3 draws
2/ Naka won't get more than 2 wins
It seems that you expect a close match, as do I. How much time (or how many nodes) did you give Lc0 in the simulation? Did you base it on Hikaru's standard FIDE rating or on his Rapid FIDE rating (which is much higher), or something else? How about the hardware for Komodo, is this just assuming we ran on your hardware? When I last tried this, Lc0 was not a good simulator of a human playing handicap, but that was a while ago, perhaps now it is okay. There's also the question of MCTS vs. normal Dragon; probably MCTS will do worse in simulation, but perhaps better against actual human.
I used vague extrapolations, but here a factor of 2, especially in Dragon speed or TC, is hardly changing anything. I took Naka's rating as 2800 FIDE Rapid, and 300-400k SF12 nodes per move needed to match Naka (2800) based on our older 45min + 30s estimates with Komodo and SF (older). At 15min + 10s this would mean at most 150k SF12 nodes per move and that is about 100-200 nodes of Lc0 (large net J92-330) as strength goes (direct match). A factor of 3-4 with Lc0 nodes per move (human TC) is Elo-wise a much larger factor of about 20-40 (from scaling properties, diminishing returns) for Dragon full speed or nodes per move at 15min + 10s. All in all I have reduced the match to the simulation of 50 nodes per move Lc0 (Drawscore=80) and 100s + 1s full strength 4 core Dragon (Contempt=100). While Dragon can be 2-3 time faster or slower without perturbing much the result, the difference with Lc0 at 30 and 100 nodes per move is not negligible, and that's pretty much my worst case and best case results for Naka. And the results show some at least a half draws (many failed Naka wins) and somewhat on average larger Dragon number of wins, but with serious possibility that Naka might win the match.
That is funny. We each had a calabration method, but came to the same conclusion in terms of Lc0 nodes per move. I just used nodes per second limit option set at 2.3. Resulting in the about the same nodes = Naka per move. And both showing a Naka win.
Yes, we had very similar conditions, after all.
My median is somewhere that Dragon is favored a bit, but it could go either way:
Jouni wrote: ↑Tue Nov 17, 2020 3:28 pm
My personal feeling: human computer era ended around 2005 when SMP Rybka was released. Handicap matches are not interesting.
I would like to see non handicap games against GM's even if the engines win every game.
Note that I am not sure that they can win every game with black and maybe humans can find some strategy that give them practical chances for a draw.
I agree. To me, the only thing a handicap match accomplishes is embarrassment to the man if he still loses.
GM Nakamura does not care about such things. He likes the competition and more views to his channel.
Here is GM Nakamura playing Komodo in a 10 game match straight up. And gets crushed 10-0.
It is interesting to note in the video in the last game. GM Nakamura says he thinks a 2 pawn advantage for him. Would be a even match setup.
Yes, and many draws. I am getting in simulations Dragon vs LC0 J92-330 "the worst case scenario" for Dragon as 2 losses and 6 draws. And the best case as 3 wins and 5 draws.
My bets would be:
1/ There will be more than 3 draws
2/ Naka won't get more than 2 wins
It seems that you expect a close match, as do I. How much time (or how many nodes) did you give Lc0 in the simulation? Did you base it on Hikaru's standard FIDE rating or on his Rapid FIDE rating (which is much higher), or something else? How about the hardware for Komodo, is this just assuming we ran on your hardware? When I last tried this, Lc0 was not a good simulator of a human playing handicap, but that was a while ago, perhaps now it is okay. There's also the question of MCTS vs. normal Dragon; probably MCTS will do worse in simulation, but perhaps better against actual human.
I used vague extrapolations, but here a factor of 2, especially in Dragon speed or TC, is hardly changing anything. I took Naka's rating as 2800 FIDE Rapid, and 300-400k SF12 nodes per move needed to match Naka (2800) based on our older 45min + 30s estimates with Komodo and SF (older). At 15min + 10s this would mean at most 150k SF12 nodes per move and that is about 100-200 nodes of Lc0 (large net J92-330) as strength goes (direct match). A factor of 3-4 with Lc0 nodes per move (human TC) is Elo-wise a much larger factor of about 20-40 (from scaling properties, diminishing returns) for Dragon full speed or nodes per move at 15min + 10s. All in all I have reduced the match to the simulation of 50 nodes per move Lc0 (Drawscore=80) and 100s + 1s full strength 4 core Dragon (Contempt=100). While Dragon can be 2-3 time faster or slower without perturbing much the result, the difference with Lc0 at 30 and 100 nodes per move is not negligible, and that's pretty much my worst case and best case results for Naka. And the results show some at least a half draws (many failed Naka wins) and somewhat on average larger Dragon number of wins, but with serious possibility that Naka might win the match.
That is funny. We each had a calabration method, but came to the same conclusion in terms of Lc0 nodes per move. I just used nodes per second limit option set at 2.3. Resulting in the about the same nodes = Naka per move. And both showing a Naka win.
Yes, we had very similar conditions, after all.
My median is somewhere that Dragon is favored a bit, but it could go either way:
Jouni wrote: ↑Tue Nov 17, 2020 3:28 pm
My personal feeling: human computer era ended around 2005 when SMP Rybka was released. Handicap matches are not interesting.
I would like to see non handicap games against GM's even if the engines win every game.
Note that I am not sure that they can win every game with black and maybe humans can find some strategy that give them practical chances for a draw.
I agree. To me, the only thing a handicap match accomplishes is embarrassment to the man if he still loses.
GM Nakamura does not care about such things. He likes the competition and more views to his channel.
Here is GM Nakamura playing Komodo in a 10 game match straight up. And gets crushed 10-0.
It is interesting to note in the video in the last game. GM Nakamura says he thinks a 2 pawn advantage for him. Would be a even match setup.
Yes, and many draws. I am getting in simulations Dragon vs LC0 J92-330 "the worst case scenario" for Dragon as 2 losses and 6 draws. And the best case as 3 wins and 5 draws.
My bets would be:
1/ There will be more than 3 draws
2/ Naka won't get more than 2 wins
It seems that you expect a close match, as do I. How much time (or how many nodes) did you give Lc0 in the simulation? Did you base it on Hikaru's standard FIDE rating or on his Rapid FIDE rating (which is much higher), or something else? How about the hardware for Komodo, is this just assuming we ran on your hardware? When I last tried this, Lc0 was not a good simulator of a human playing handicap, but that was a while ago, perhaps now it is okay. There's also the question of MCTS vs. normal Dragon; probably MCTS will do worse in simulation, but perhaps better against actual human.
I used vague extrapolations, but here a factor of 2, especially in Dragon speed or TC, is hardly changing anything. I took Naka's rating as 2800 FIDE Rapid, and 300-400k SF12 nodes per move needed to match Naka (2800) based on our older 45min + 30s estimates with Komodo and SF (older). At 15min + 10s this would mean at most 150k SF12 nodes per move and that is about 100-200 nodes of Lc0 (large net J92-330) as strength goes (direct match). A factor of 3-4 with Lc0 nodes per move (human TC) is Elo-wise a much larger factor of about 20-40 (from scaling properties, diminishing returns) for Dragon full speed or nodes per move at 15min + 10s. All in all I have reduced the match to the simulation of 50 nodes per move Lc0 (Drawscore=80) and 100s + 1s full strength 4 core Dragon (Contempt=100). While Dragon can be 2-3 time faster or slower without perturbing much the result, the difference with Lc0 at 30 and 100 nodes per move is not negligible, and that's pretty much my worst case and best case results for Naka. And the results show some at least a half draws (many failed Naka wins) and somewhat on average larger Dragon number of wins, but with serious possibility that Naka might win the match.
That is funny. We each had a calabration method, but came to the same conclusion in terms of Lc0 nodes per move. I just used nodes per second limit option set at 2.3. Resulting in the about the same nodes = Naka per move. And both showing a Naka win.
Yes, we had very similar conditions, after all.
My median is somewhere that Dragon is favored a bit, but it could go either way:
Jouni wrote: ↑Tue Nov 17, 2020 3:28 pm
My personal feeling: human computer era ended around 2005 when SMP Rybka was released. Handicap matches are not interesting.
I would like to see non handicap games against GM's even if the engines win every game.
Note that I am not sure that they can win every game with black and maybe humans can find some strategy that give them practical chances for a draw.
I agree. To me, the only thing a handicap match accomplishes is embarrassment to the man if he still loses.
GM Nakamura does not care about such things. He likes the competition and more views to his channel.
Here is GM Nakamura playing Komodo in a 10 game match straight up. And gets crushed 10-0.
It is interesting to note in the video in the last game. GM Nakamura says he thinks a 2 pawn advantage for him. Would be a even match setup.
Yes, and many draws. I am getting in simulations Dragon vs LC0 J92-330 "the worst case scenario" for Dragon as 2 losses and 6 draws. And the best case as 3 wins and 5 draws.
My bets would be:
1/ There will be more than 3 draws
2/ Naka won't get more than 2 wins
It seems that you expect a close match, as do I. How much time (or how many nodes) did you give Lc0 in the simulation? Did you base it on Hikaru's standard FIDE rating or on his Rapid FIDE rating (which is much higher), or something else? How about the hardware for Komodo, is this just assuming we ran on your hardware? When I last tried this, Lc0 was not a good simulator of a human playing handicap, but that was a while ago, perhaps now it is okay. There's also the question of MCTS vs. normal Dragon; probably MCTS will do worse in simulation, but perhaps better against actual human.
I used vague extrapolations, but here a factor of 2, especially in Dragon speed or TC, is hardly changing anything. I took Naka's rating as 2800 FIDE Rapid, and 300-400k SF12 nodes per move needed to match Naka (2800) based on our older 45min + 30s estimates with Komodo and SF (older). At 15min + 10s this would mean at most 150k SF12 nodes per move and that is about 100-200 nodes of Lc0 (large net J92-330) as strength goes (direct match). A factor of 3-4 with Lc0 nodes per move (human TC) is Elo-wise a much larger factor of about 20-40 (from scaling properties, diminishing returns) for Dragon full speed or nodes per move at 15min + 10s. All in all I have reduced the match to the simulation of 50 nodes per move Lc0 (Drawscore=80) and 100s + 1s full strength 4 core Dragon (Contempt=100). While Dragon can be 2-3 time faster or slower without perturbing much the result, the difference with Lc0 at 30 and 100 nodes per move is not negligible, and that's pretty much my worst case and best case results for Naka. And the results show some at least a half draws (many failed Naka wins) and somewhat on average larger Dragon number of wins, but with serious possibility that Naka might win the match.
That is funny. We each had a calabration method, but came to the same conclusion in terms of Lc0 nodes per move. I just used nodes per second limit option set at 2.3. Resulting in the about the same nodes = Naka per move. And both showing a Naka win.
Yes, we had very similar conditions, after all.
My median is somewhere that Dragon is favored a bit, but it could go either way:
Score of Dragon_C100 vs Lc0_J92_330_CUDA: 6 - 3 - 23 [0.547] 32
... Dragon_C100 playing White: 6 - 3 - 23 [0.547] 32
... White vs Black: 6 - 3 - 23 [0.547] 32
Elo difference: 32.7 +/- 64.0, LOS: 84.1 %, DrawRatio: 71.9 %
Finished match
So the result after 32 games would predict:
Naka 0.75 wins
Dragon 1.5 wins
and 5.75 draws
And the range of outcomes, as I showed earlier, seems the following:
Dragon: 0 to 3 wins
Naka: 0 to 2 wins
and 4 to 7 draws
You should be pretty happy! Sim Nakamura results holding up in day one.
Hikaru, might have performed better if he were not chatting with the viewers at the same time as playing Komodo. These machines require your undivided attention.
Jouni wrote: ↑Tue Nov 17, 2020 3:28 pm
My personal feeling: human computer era ended around 2005 when SMP Rybka was released. Handicap matches are not interesting.
I would like to see non handicap games against GM's even if the engines win every game.
Note that I am not sure that they can win every game with black and maybe humans can find some strategy that give them practical chances for a draw.
I agree. To me, the only thing a handicap match accomplishes is embarrassment to the man if he still loses.
GM Nakamura does not care about such things. He likes the competition and more views to his channel.
Here is GM Nakamura playing Komodo in a 10 game match straight up. And gets crushed 10-0.
It is interesting to note in the video in the last game. GM Nakamura says he thinks a 2 pawn advantage for him. Would be a even match setup.
Yes, and many draws. I am getting in simulations Dragon vs LC0 J92-330 "the worst case scenario" for Dragon as 2 losses and 6 draws. And the best case as 3 wins and 5 draws.
My bets would be:
1/ There will be more than 3 draws
2/ Naka won't get more than 2 wins
It seems that you expect a close match, as do I. How much time (or how many nodes) did you give Lc0 in the simulation? Did you base it on Hikaru's standard FIDE rating or on his Rapid FIDE rating (which is much higher), or something else? How about the hardware for Komodo, is this just assuming we ran on your hardware? When I last tried this, Lc0 was not a good simulator of a human playing handicap, but that was a while ago, perhaps now it is okay. There's also the question of MCTS vs. normal Dragon; probably MCTS will do worse in simulation, but perhaps better against actual human.
I used vague extrapolations, but here a factor of 2, especially in Dragon speed or TC, is hardly changing anything. I took Naka's rating as 2800 FIDE Rapid, and 300-400k SF12 nodes per move needed to match Naka (2800) based on our older 45min + 30s estimates with Komodo and SF (older). At 15min + 10s this would mean at most 150k SF12 nodes per move and that is about 100-200 nodes of Lc0 (large net J92-330) as strength goes (direct match). A factor of 3-4 with Lc0 nodes per move (human TC) is Elo-wise a much larger factor of about 20-40 (from scaling properties, diminishing returns) for Dragon full speed or nodes per move at 15min + 10s. All in all I have reduced the match to the simulation of 50 nodes per move Lc0 (Drawscore=80) and 100s + 1s full strength 4 core Dragon (Contempt=100). While Dragon can be 2-3 time faster or slower without perturbing much the result, the difference with Lc0 at 30 and 100 nodes per move is not negligible, and that's pretty much my worst case and best case results for Naka. And the results show some at least a half draws (many failed Naka wins) and somewhat on average larger Dragon number of wins, but with serious possibility that Naka might win the match.
That is funny. We each had a calabration method, but came to the same conclusion in terms of Lc0 nodes per move. I just used nodes per second limit option set at 2.3. Resulting in the about the same nodes = Naka per move. And both showing a Naka win.
Yes, we had very similar conditions, after all.
My median is somewhere that Dragon is favored a bit, but it could go either way:
Score of Dragon_C100 vs Lc0_J92_330_CUDA: 6 - 3 - 23 [0.547] 32
... Dragon_C100 playing White: 6 - 3 - 23 [0.547] 32
... White vs Black: 6 - 3 - 23 [0.547] 32
Elo difference: 32.7 +/- 64.0, LOS: 84.1 %, DrawRatio: 71.9 %
Finished match
So the result after 32 games would predict:
Naka 0.75 wins
Dragon 1.5 wins
and 5.75 draws
And the range of outcomes, as I showed earlier, seems the following:
Dragon: 0 to 3 wins
Naka: 0 to 2 wins
and 4 to 7 draws
You should be pretty happy! Sim Nakamura results holding up in day one.
Hikaru, might have performed better if he were not chatting with the viewers at the same time as playing Komodo. These machines require your undivided attention.
I totally agree with you, my prediction of at least three wins for Naka and 5 wins was based on thinking that Komodo Dragon was going to be switching White and Black side with the same Odds. Jorge Sammour with a Knight Odds would probably win at least 3 games and draw 2 using my 8 cores system. I only watched the first three games what was the final result ?
