Dragon vs GM Nakamura Analog Handicap Match

Discussion of anything and everything relating to chess playing software and machines.

Moderators: Harvey Williamson, Dann Corbit, hgm

Forum rules
This textbox is used to restore diagrams posted with the [d] tag before the upgrade.
lkaufman
Posts: 4426
Joined: Sun Jan 10, 2010 5:15 am
Location: Maryland USA
Contact:

Re: Dragon vs GM Nakamura Analog Handicap Match

Post by lkaufman » Thu Nov 19, 2020 2:57 am

Nordlandia wrote:
Thu Nov 19, 2020 12:15 am
The computer has to play the white pieces ?. The question is whether Two Pawns + Move is too much against Nakamura !?
Dragon played a bunch of games at this time control with a kid rated about 2500 on chess.com blitz at two pawns and move (the same four pairs of pawns as in this match). Dragon came out ahead, but not way ahead. So maybe it's about right for a relatively low rated GM, maybe rated about 2600 or so on chess.com blitz; Nakamura is about 3200 on chess.com blitz, so he would win pretty easily. Even I do pretty well when the two pawns are c7 + f7, which is really a big handicap, though still less than knight odds.
Komodo rules!

Chessqueen
Posts: 1129
Joined: Wed Sep 05, 2018 12:16 am
Full name: Nancy M Pichardo

Re: Dragon vs GM Nakamura Analog Handicap Match

Post by Chessqueen » Thu Nov 19, 2020 3:58 am

lkaufman wrote:
Thu Nov 19, 2020 2:57 am
Nordlandia wrote:
Thu Nov 19, 2020 12:15 am
The computer has to play the white pieces ?. The question is whether Two Pawns + Move is too much against Nakamura !?
Dragon played a bunch of games at this time control with a kid rated about 2500 on chess.com blitz at two pawns and move (the same four pairs of pawns as in this match). Dragon came out ahead, but not way ahead. So maybe it's about right for a relatively low rated GM, maybe rated about 2600 or so on chess.com blitz; Nakamura is about 3200 on chess.com blitz, so he would win pretty easily. Even I do pretty well when the two pawns are c7 + f7, which is really a big handicap, though still less than knight odds.
What was the final score GM Naka Vs Komodo Dragon >

Chessqueen
Posts: 1129
Joined: Wed Sep 05, 2018 12:16 am
Full name: Nancy M Pichardo

Re: Dragon vs GM Nakamura Analog Handicap Match

Post by Chessqueen » Thu Nov 19, 2020 4:17 am

mwyoung wrote:
Wed Nov 18, 2020 8:11 pm
lkaufman wrote:
Wed Nov 18, 2020 7:22 pm
mwyoung wrote:
Wed Nov 18, 2020 7:04 pm
mwyoung wrote:
Wed Nov 18, 2020 12:38 pm
lkaufman wrote:
Tue Nov 17, 2020 8:22 pm
mwyoung wrote:
Tue Nov 17, 2020 7:17 pm
lkaufman wrote:
Tue Nov 17, 2020 4:34 pm
What I've consistently observed is that humans always do better taking handicaps from top engines than similarly rated engines do. Maybe it's because they have a better understanding of simplifying when ahead, or just that they know to respect the opponent and avoid anything unclear. I have been searching for a computer opponent that can truly simulate a human GM in these handicap matches, but I haven't really found one yet. It's easy for me to run such simulations, but they don't predict well how the human will do.
So you want more brains and less calculation. Here is what I came up with to mimic a result. Lc0-CPU (J94-40) and set the nodes per second option = 4. Lc0 is searching to only 3 or 4 ply. But does much better then the A/B engines with almost now search. Is the style more GM like? But it is fun to see what it takes to win this match for the human player.

Code: Select all

DESKTOP-CORSAIR, Blitz 5.0min+1.0sec  0

                                         1234
1   Lc0 v0.26.3 - CPU               +89   1½½½   2.5/4
2   Dragon by Komodo Chess 64-bit   -89  0½½½    1.5/4






To complete the experiment you should run the crippled Lc0 against the ancient SF in standard chess to see if they are really of roughly comparable strength. If they are then this might indeed be a good way to simulate a human.
I calibrated my simulated human player better.

With Lc0-CPU and NN (J94-40) At rapid time controls. The base Elo was 2300 Elo ~. I can not go lower with this net strength.

Calibration settings Lc0 (J94-40) with Node per second option = 0.1 Elo = 2300~. Each 0.1 increased gained 25 Elo ~.

Results playing Stockfish 1.4 In a even match against a simulated Nakamura. Naka lost badly by over 200 Elo. And this is what you would expect.

Results playing Dragon with a 2 pawn odds match. Where Stockfish 1.4 lost all games. Simulated Nakamura does much better. Simulation still running.

Simulation shows a 2300 Elo player should beat Dragon at knight odds. I know your son is lower rated then 2300 and had a even match in 4 games.

When the simulation is complete. I will give the simulated results of the upcoming Dragon vs Naka odds match. And we can see if it can predict the winner.
The simulation is completed. It shows simulated Nakamura using Naka's rapid rating winning the match against my machine running Dragon. Your machine is faster, but we do not know how well Nakamura will prep for the opening positions. Given that I will have to go with GM Nakamura still winning the match.

Code: Select all

DESKTOP-CORSAIR, Rapid 15.0min+10.0sec  0

                                         12345678
1   Lc0 v0.26.3 - CPU               +44   ½½½11½0½   4.5/8
2   Dragon by Komodo Chess 64-bit   -44  ½½½00½1½    3.5/8

For fun and a sanity check with my simulated Nakamura. I also run the match at 3m+0s. Using GM Nakamura's Blitz rating in my simulation. This put Nakamura's simulation setting at Node Per Second = 2.5. Here is that match results.

Code: Select all

DESKTOP-CORSAIR, Blitz 3.0min  0

                                         12345678
1   Dragon by Komodo Chess 64-bit  +255   1½1½1½11   6.5/8
2   Lc0 v0.26.3 - CPU              -255  0½0½0½00    1.5/8

Simuation games of 15m + 10s odds match.




It does not. I had to do the calibration work. This is my own concept, but it looks like Lasko had a simular concept.

I just took Lc0 - cpu version. And a big net like J94-40. I think Lasko used J92-330. I determained the rating with test games with the option nodes per second limit set at 0.1. This gave me my base rating with net J94-40 at 2300 Elo. Then with more test games. I determained how increasing the node per second limit option affects performance. With J94-40 each 0.1 increase in nodes per second increased performance by about 25 Elo.

Now with this information you can calibrate your sim human by rating. And the sim human now plays tactically better with more time, worse with less time. But still keeping the positional knowledge the same. So it plays more realistic in odds match games.

Does this sim human work? This will depend on if the predicted results are correct.
As you know NOW programs do NOT play like Humans, and most programs do NOT force to trade pieces when they are ahead nor close the position to force the program into time problem causing it to make a desperate and blundering move

mwyoung
Posts: 2577
Joined: Wed May 12, 2010 8:00 pm

Re: Dragon vs GM Nakamura Analog Handicap Match

Post by mwyoung » Thu Nov 19, 2020 4:24 am

Chessqueen wrote:
Thu Nov 19, 2020 4:17 am
mwyoung wrote:
Wed Nov 18, 2020 8:11 pm
lkaufman wrote:
Wed Nov 18, 2020 7:22 pm
mwyoung wrote:
Wed Nov 18, 2020 7:04 pm
mwyoung wrote:
Wed Nov 18, 2020 12:38 pm
lkaufman wrote:
Tue Nov 17, 2020 8:22 pm
mwyoung wrote:
Tue Nov 17, 2020 7:17 pm
lkaufman wrote:
Tue Nov 17, 2020 4:34 pm
What I've consistently observed is that humans always do better taking handicaps from top engines than similarly rated engines do. Maybe it's because they have a better understanding of simplifying when ahead, or just that they know to respect the opponent and avoid anything unclear. I have been searching for a computer opponent that can truly simulate a human GM in these handicap matches, but I haven't really found one yet. It's easy for me to run such simulations, but they don't predict well how the human will do.
So you want more brains and less calculation. Here is what I came up with to mimic a result. Lc0-CPU (J94-40) and set the nodes per second option = 4. Lc0 is searching to only 3 or 4 ply. But does much better then the A/B engines with almost now search. Is the style more GM like? But it is fun to see what it takes to win this match for the human player.

Code: Select all

DESKTOP-CORSAIR, Blitz 5.0min+1.0sec  0

                                         1234
1   Lc0 v0.26.3 - CPU               +89   1½½½   2.5/4
2   Dragon by Komodo Chess 64-bit   -89  0½½½    1.5/4






To complete the experiment you should run the crippled Lc0 against the ancient SF in standard chess to see if they are really of roughly comparable strength. If they are then this might indeed be a good way to simulate a human.
I calibrated my simulated human player better.

With Lc0-CPU and NN (J94-40) At rapid time controls. The base Elo was 2300 Elo ~. I can not go lower with this net strength.

Calibration settings Lc0 (J94-40) with Node per second option = 0.1 Elo = 2300~. Each 0.1 increased gained 25 Elo ~.

Results playing Stockfish 1.4 In a even match against a simulated Nakamura. Naka lost badly by over 200 Elo. And this is what you would expect.

Results playing Dragon with a 2 pawn odds match. Where Stockfish 1.4 lost all games. Simulated Nakamura does much better. Simulation still running.

Simulation shows a 2300 Elo player should beat Dragon at knight odds. I know your son is lower rated then 2300 and had a even match in 4 games.

When the simulation is complete. I will give the simulated results of the upcoming Dragon vs Naka odds match. And we can see if it can predict the winner.
The simulation is completed. It shows simulated Nakamura using Naka's rapid rating winning the match against my machine running Dragon. Your machine is faster, but we do not know how well Nakamura will prep for the opening positions. Given that I will have to go with GM Nakamura still winning the match.

Code: Select all

DESKTOP-CORSAIR, Rapid 15.0min+10.0sec  0

                                         12345678
1   Lc0 v0.26.3 - CPU               +44   ½½½11½0½   4.5/8
2   Dragon by Komodo Chess 64-bit   -44  ½½½00½1½    3.5/8

For fun and a sanity check with my simulated Nakamura. I also run the match at 3m+0s. Using GM Nakamura's Blitz rating in my simulation. This put Nakamura's simulation setting at Node Per Second = 2.5. Here is that match results.

Code: Select all

DESKTOP-CORSAIR, Blitz 3.0min  0

                                         12345678
1   Dragon by Komodo Chess 64-bit  +255   1½1½1½11   6.5/8
2   Lc0 v0.26.3 - CPU              -255  0½0½0½00    1.5/8

Simuation games of 15m + 10s odds match.




It does not. I had to do the calibration work. This is my own concept, but it looks like Lasko had a simular concept.

I just took Lc0 - cpu version. And a big net like J94-40. I think Lasko used J92-330. I determained the rating with test games with the option nodes per second limit set at 0.1. This gave me my base rating with net J94-40 at 2300 Elo. Then with more test games. I determained how increasing the node per second limit option affects performance. With J94-40 each 0.1 increase in nodes per second increased performance by about 25 Elo.

Now with this information you can calibrate your sim human by rating. And the sim human now plays tactically better with more time, worse with less time. But still keeping the positional knowledge the same. So it plays more realistic in odds match games.

Does this sim human work? This will depend on if the predicted results are correct.
As you know NOW programs do NOT play like Humans, and most programs do NOT force to trade pieces when they are ahead nor close the position to force the program into time problem causing it to make a desperate and blundering move
Strang response. Since the predicted stats held up day 1. You could be totally right tomorrow, but not today.
"The worst thing that can happen to a forum is a running wild attacking moderator(HGM) who is not corrected by the community." - Ed Schröder

Chessqueen
Posts: 1129
Joined: Wed Sep 05, 2018 12:16 am
Full name: Nancy M Pichardo

Re: Dragon vs GM Nakamura Analog Handicap Match

Post by Chessqueen » Thu Nov 19, 2020 3:52 pm

lkaufman wrote:
Wed Nov 18, 2020 12:26 am
Dragon vs. my son Ray was drawn 2 to 2 at knight odds, same tc as the Naka match. So if the difference in the odds offsets the strength difference between Naka and Ray, the Dragon vs Naka match should be close. It should be noted that standard Dragon lost 1.5 to 0.5 to Ray, while MCTS Dragon won by 1.5 to 0.5. Of course that's not a statistically significant difference, just a hint that it's okay to use MCTS for at least some games.
Can you post those 4 games

lkaufman
Posts: 4426
Joined: Sun Jan 10, 2010 5:15 am
Location: Maryland USA
Contact:

Re: Dragon vs GM Nakamura Analog Handicap Match

Post by lkaufman » Thu Nov 19, 2020 4:09 pm

Chessqueen wrote:
Thu Nov 19, 2020 3:52 pm
lkaufman wrote:
Wed Nov 18, 2020 12:26 am
Dragon vs. my son Ray was drawn 2 to 2 at knight odds, same tc as the Naka match. So if the difference in the odds offsets the strength difference between Naka and Ray, the Dragon vs Naka match should be close. It should be noted that standard Dragon lost 1.5 to 0.5 to Ray, while MCTS Dragon won by 1.5 to 0.5. Of course that's not a statistically significant difference, just a hint that it's okay to use MCTS for at least some games.
Can you post those 4 games
Links to them with annotations were posted in the responses to the announcement of the match on chess.com.
Komodo rules!

User avatar
Laskos
Posts: 10945
Joined: Wed Jul 26, 2006 8:21 pm
Full name: Kai Laskos

Re: Dragon vs GM Nakamura Analog Handicap Match

Post by Laskos » Thu Nov 19, 2020 9:02 pm

lkaufman wrote:
Thu Nov 19, 2020 2:57 am
Nordlandia wrote:
Thu Nov 19, 2020 12:15 am
The computer has to play the white pieces ?. The question is whether Two Pawns + Move is too much against Nakamura !?
Dragon played a bunch of games at this time control with a kid rated about 2500 on chess.com blitz at two pawns and move (the same four pairs of pawns as in this match). Dragon came out ahead, but not way ahead. So maybe it's about right for a relatively low rated GM, maybe rated about 2600 or so on chess.com blitz; Nakamura is about 3200 on chess.com blitz, so he would win pretty easily. Even I do pretty well when the two pawns are c7 + f7, which is really a big handicap, though still less than knight odds.

Are you sure it would be so easy to Naka? My simulations show some +2 -1 =5 expected outcome for Naka (32 games check), and still difficulties of converting the wins. But the simulations are very unreliable, and the openings here might be more decisive, as it is the human moving first into the weak Black defense.

Chessqueen
Posts: 1129
Joined: Wed Sep 05, 2018 12:16 am
Full name: Nancy M Pichardo

Re: Dragon vs GM Nakamura Analog Handicap Match

Post by Chessqueen » Thu Nov 19, 2020 9:26 pm

Laskos wrote:
Thu Nov 19, 2020 9:02 pm
lkaufman wrote:
Thu Nov 19, 2020 2:57 am
Nordlandia wrote:
Thu Nov 19, 2020 12:15 am
The computer has to play the white pieces ?. The question is whether Two Pawns + Move is too much against Nakamura !?
Dragon played a bunch of games at this time control with a kid rated about 2500 on chess.com blitz at two pawns and move (the same four pairs of pawns as in this match). Dragon came out ahead, but not way ahead. So maybe it's about right for a relatively low rated GM, maybe rated about 2600 or so on chess.com blitz; Nakamura is about 3200 on chess.com blitz, so he would win pretty easily. Even I do pretty well when the two pawns are c7 + f7, which is really a big handicap, though still less than knight odds.

Are you sure it would be so easy to Naka? My simulations show some +2 -1 =5 expected outcome for Naka (32 games check), and still difficulties of converting the wins. But the simulations are very unreliable, and the openings here might be more decisive, as it is the human moving first into the weak Black defense.
Nakamura is Playing versus Fat Fritz he will be exhausted or tired by the time he play versus Komodo Dragon LIVE https://www.twitch.tv/gmhikaru

mehmet123
Posts: 261
Joined: Sun Jan 26, 2020 9:38 pm
Location: Turkey
Full name: Mehmet Karaman

Re: Dragon vs GM Nakamura Analog Handicap Match

Post by mehmet123 » Thu Nov 19, 2020 9:38 pm

Chessqueen wrote:
Thu Nov 19, 2020 9:26 pm

Nakamura is Playing versus Fat Fritz he will be exhausted or tired by the time he play versus Komodo Dragon LIVE https://www.twitch.tv/gmhikaru
He isn' t playing versus Fat Fritz. He is watching and commenting Chess.com Computer Chess Championship matches.

Chessqueen
Posts: 1129
Joined: Wed Sep 05, 2018 12:16 am
Full name: Nancy M Pichardo

Re: Dragon vs GM Nakamura Analog Handicap Match

Post by Chessqueen » Thu Nov 19, 2020 9:42 pm

mehmet123 wrote:
Thu Nov 19, 2020 9:38 pm
Chessqueen wrote:
Thu Nov 19, 2020 9:26 pm

Nakamura is Playing versus Fat Fritz he will be exhausted or tired by the time he play versus Komodo Dragon LIVE https://www.twitch.tv/gmhikaru
He isn' t playing versus Fat Fritz. He is watching and commenting Chess.com Computer Chess Championship matches.
Thanks for the correction, I realized it after I posted it, but his brain will be exhausted by watching and commenting. I wonder how Nakamura would do versus Fat Fritz same Odds :?: https://www.twitch.tv/gmhikaru

I believe that is Wrong to List Komodo as 3400 when it is playing with two pawns Odds it is probably very close to Nakamura rating under this condition. People are taking his concentration away by asking NAKA STUPID question, so he is playing with a BIG handicap https://www.twitch.tv/gmhikaru

Post Reply