Knight odds test blitz match dragon vs GM Lenderman

[lkaufman]

I truly believe that after you give a Knight odds to any GM rated above 2675, once he realized that the position is equal having the same or equal amount of time left, once Komodo Dragon level the position there is no reason to continue against an engine rated over 650 rating points above the GM

CCRL rating != FIDE rating

Even if you reduce the CCRL by 200 rating points Dragon is still 550 rating more, remember when Deep Fritz and Deep Junior was only about 250 CCRL rating point over GM Kramnik, and when they played a match with even pieces, Well you know what happened.
https://www.nytimes.com/2006/12/05/cros ... chess.html

https://en.chessbase.com/post/kramnik-v ... tch-by-4-2

Actually 550 or 650 are gross underestimates of the elo difference if we're talking about blitz, which is what is relevant here. If we go by the CCRL scale, Dragon on 8 cores is 3679, and the latest Dragon in MCTS mode (in Linux so not affected by the Windows MCTS compile issue reported in another thread) on 32 cores should be around 3700 CCRL blitz (quadruple cores should more than offset MCTS). Based on his games with crippled Komodo versions, I think Nakamura would be something like 2250 against the engines under CCRL blitz conditions, and he is about 250 above Lenderman on FIDE blitz, so Lenderman should be 2000 CCRL blitz (!). That's 1700 difference! Or if you use the chess.com blitz scale, Lenderman is a bit under 2800, Nakamura usually near 3200. Komodo (pre-dragon, using 16 of 32 cores) beat him 10 to zero at ten minute chess, and beat him 6.5 to 1.5 giving him two pawns in rapid; I think Dragon would be over 4000 against the top humans playing chess.com blitz. That's a 1200 gap. Maybe that's more realistic than 1700 (the human scale is more compact than the engine scale), but anyway a lot more than 550 or 650. In general my testing has indicated that knight odds is somewhere in the 1000 to 1200 elo range at GM level for the odds-receiver, so this is pretty consistent with the match result. At classical time limits I think that the CCRL 40/15 list is not too far off from FIDE, where Dragon on 4 cpu is about 3500, so in classical chess a 2400 FIDE IM should be a good match based on this. Actually I think the IM would win, the elo gap for knight odds is greater with longer time controls.

Thank Mr. Kaufman, I totally agree with you, except " I think Nakamura would be something like 2250 against the engines under CCRL blitz conditions" That would only apply to a 2250 CCRL rated engine that uses good chess opening, Plus at least a 4 pieces EGTB, but if you take for instance Big Jump rated around 2250 without a good chess opening and EGTB I would bet on GM Nakamura playing 5 Minutes Blitz.

I don't see a "Big Jump" on the list, but you probably mean BikJump which is 2102 on CCRL blitz, so Nakamura would have to score about 70% against it on an I7 in blitz, using at least a decent opening book, just to get a 2250 rating. That seems realistic to me based on his play vs. the chess.com bots. Also note that "CCRL blitz conditions" means 2' +1", so roughly 3 min chess rather than 5 min chess. Maybe at 5 min he would be 2300 CCRL blitz.

Here we clearly see GM Nakamura beating all the chess.com bots

Those are not all the chess.com bots, they seem to stop at elo 2100. On the show I watched, he played bots in the 2500 to 2750 elo range and generally had a tough time. He scored somewhere around even with the ones in the 2500s, but when he had to play "himself" (the Hikaru bot), he lost repeatedly, I don't remember if he finally won a game or not. These bots would be several hundred elo lower on CCRL blitz scale.

[Chessqueen]

I truly believe that after you give a Knight odds to any GM rated above 2675, once he realized that the position is equal having the same or equal amount of time left, once Komodo Dragon level the position there is no reason to continue against an engine rated over 650 rating points above the GM

CCRL rating != FIDE rating

Even if you reduce the CCRL by 200 rating points Dragon is still 550 rating more, remember when Deep Fritz and Deep Junior was only about 250 CCRL rating point over GM Kramnik, and when they played a match with even pieces, Well you know what happened.
https://www.nytimes.com/2006/12/05/cros ... chess.html

https://en.chessbase.com/post/kramnik-v ... tch-by-4-2

Actually 550 or 650 are gross underestimates of the elo difference if we're talking about blitz, which is what is relevant here. If we go by the CCRL scale, Dragon on 8 cores is 3679, and the latest Dragon in MCTS mode (in Linux so not affected by the Windows MCTS compile issue reported in another thread) on 32 cores should be around 3700 CCRL blitz (quadruple cores should more than offset MCTS). Based on his games with crippled Komodo versions, I think Nakamura would be something like 2250 against the engines under CCRL blitz conditions, and he is about 250 above Lenderman on FIDE blitz, so Lenderman should be 2000 CCRL blitz (!). That's 1700 difference! Or if you use the chess.com blitz scale, Lenderman is a bit under 2800, Nakamura usually near 3200. Komodo (pre-dragon, using 16 of 32 cores) beat him 10 to zero at ten minute chess, and beat him 6.5 to 1.5 giving him two pawns in rapid; I think Dragon would be over 4000 against the top humans playing chess.com blitz. That's a 1200 gap. Maybe that's more realistic than 1700 (the human scale is more compact than the engine scale), but anyway a lot more than 550 or 650. In general my testing has indicated that knight odds is somewhere in the 1000 to 1200 elo range at GM level for the odds-receiver, so this is pretty consistent with the match result. At classical time limits I think that the CCRL 40/15 list is not too far off from FIDE, where Dragon on 4 cpu is about 3500, so in classical chess a 2400 FIDE IM should be a good match based on this. Actually I think the IM would win, the elo gap for knight odds is greater with longer time controls.

Thank Mr. Kaufman, I totally agree with you, except " I think Nakamura would be something like 2250 against the engines under CCRL blitz conditions" That would only apply to a 2250 CCRL rated engine that uses good chess opening, Plus at least a 4 pieces EGTB, but if you take for instance Big Jump rated around 2250 without a good chess opening and EGTB I would bet on GM Nakamura playing 5 Minutes Blitz.

I don't see a "Big Jump" on the list, but you probably mean BikJump which is 2102 on CCRL blitz, so Nakamura would have to score about 70% against it on an I7 in blitz, using at least a decent opening book, just to get a 2250 rating. That seems realistic to me based on his play vs. the chess.com bots. Also note that "CCRL blitz conditions" means 2' +1", so roughly 3 min chess rather than 5 min chess. Maybe at 5 min he would be 2300 CCRL blitz.

Here we clearly see GM Nakamura beating all the chess.com bots

Those are not all the chess.com bots, they seem to stop at elo 2100. On the show I watched, he played bots in the 2500 to 2750 elo range and generally had a tough time. He scored somewhere around even with the ones in the 2500s, but when he had to play "himself" (the Hikaru bot), he lost repeatedly, I don't remember if he finally won a game or not. These bots would be several hundred elo lower on CCRL blitz scale.

That is true, GM Nakamura had to tried more than 5 times in order to beat GM Maurice Bot rated around 2500. That is very funny to say the least, I know that a bot simulating to play like yourself it is not even close since human with pressure blunder and GM Nakamura will never beat a bot of GM Nakamura strength

Note: Based on your estimate Dragon playing at TC 2'+1" with Knight Odds is probably rated around 2350 or higher ? I can not wait until dragon beat GM Nakamura at Blitz 2'+1" with knight Odds

[towforce]

Some moderation penalty on useless permanent full quoting would that help?

This thread should be added to the internet-wiki, as a role model, how a troll lures another member into full quoting...

Watch the troll-barometer!

It seems unfair to criticise for full quoting unless you have a rule, or at least a guideline, for maximum nesting of quotes.

I suggest a guideline of a maximum nesting depth of 4. Much more than that quickly becomes more harmful than helpful.

Level 1

Level 2

Level 3

Level 4

[Chessqueen]

Some moderation penalty on useless permanent full quoting would that help?

This thread should be added to the internet-wiki, as a role model, how a troll lures another member into full quoting...

Watch the troll-barometer!

It seems unfair to criticise for full quoting unless you have a rule, or at least a guideline, for maximum nesting of quotes.

I suggest a guideline of a maximum nesting depth of 4. Much more than that quickly becomes more harmful than helpful.

Level 1

Level 2

Level 3

Level 4

To ask questions to one of the most capable person on computer chess is NOT trolling, unlike most of you who were NOT even born back in 1987, I have been following Mr. Kaufman and I started buying his original Magazine of Computer Chess Report since 1987, and I have the most respect for Mr. Kaufman than anybody on this Forum

[lkaufman]

Here we clearly see GM Nakamura beating all the chess.com bots

Those are not all the chess.com bots, they seem to stop at elo 2100. On the show I watched, he played bots in the 2500 to 2750 elo range and generally had a tough time. He scored somewhere around even with the ones in the 2500s, but when he had to play "himself" (the Hikaru bot), he lost repeatedly, I don't remember if he finally won a game or not. These bots would be several hundred elo lower on CCRL blitz scale.
[/quote]

That is true, GM Nakamura had to tried more than 5 times in order to beat GM Maurice Bot rated around 2500. That is very funny to say the least, I know that a bot simulating to play like yourself it is not even close since human with pressure blunder and GM Nakamura will never beat a bot of GM Nakamura strength

Note: Based on your estimate Dragon playing at TC 2'+1" with Knight Odds is probably rated around 2350 or higher ? I can not wait until dragon beat GM Nakamura at Blitz 2'+1" with knight Odds
[/quote]

I don't know what list "2350" refers to. In engine vs engine play at 2' + 1", a recent Dragon dev. version on one thread performed at 2903 CCRL blitz giving knight odds to Wasp 2 on one thread, and at 2961 with both using four threads (several hundred games each). Against humans, Dragon giving knight odds, which means playing White, recent dragon MCTS on 32 threads was clearly too strong for GM Lenderman at 2' + 1" (he needed at least 4' + 1" to score decently as Black), might be close with Nakamura, who is usually upper 3100s chess.com blitz, 2900 FIDE blitz, and perhaps something like 2300 CCRL blitz scale. Obviously there is a huge difference between giving knight odds to a human and giving knight odds to an engine, since I'm quite sure that Nakamura would get crushed playing 2' + 1" chess on even terms with Wasp 2, even on one thread. Since modern engines do understand something about trading when ahead, I'm not fully sure why this should be so, but that's the case. Humans know to "keep it simple" against a stronger opponent giving a handicap, but the engines have at least some sense of this, so it's a bit puzzling to me.

[Chessqueen]

Here we clearly see GM Nakamura beating all the chess.com bots

Those are not all the chess.com bots, they seem to stop at elo 2100. On the show I watched, he played bots in the 2500 to 2750 elo range and generally had a tough time. He scored somewhere around even with the ones in the 2500s, but when he had to play "himself" (the Hikaru bot), he lost repeatedly, I don't remember if he finally won a game or not. These bots would be several hundred elo lower on CCRL blitz scale.

That is true, GM Nakamura had to tried more than 5 times in order to beat GM Maurice Bot rated around 2500. That is very funny to say the least, I know that a bot simulating to play like yourself it is not even close since human with pressure blunder and GM Nakamura will never beat a bot of GM Nakamura strength

Note: Based on your estimate Dragon playing at TC 2'+1" with Knight Odds is probably rated around 2350 or higher ? I can not wait until dragon beat GM Nakamura at Blitz 2'+1" with knight Odds
[/quote]

I don't know what list "2350" refers to. In engine vs engine play at 2' + 1", a recent Dragon dev. version on one thread performed at 2903 CCRL blitz giving knight odds to Wasp 2 on one thread, and at 2961 with both using four threads (several hundred games each). Against humans, Dragon giving knight odds, which means playing White, recent dragon MCTS on 32 threads was clearly too strong for GM Lenderman at 2' + 1" (he needed at least 4' + 1" to score decently as Black), might be close with Nakamura, who is usually upper 3100s chess.com blitz, 2900 FIDE blitz, and perhaps something like 2300 CCRL blitz scale. Obviously there is a huge difference between giving knight odds to a human and giving knight odds to an engine, since I'm quite sure that Nakamura would get crushed playing 2' + 1" chess on even terms with Wasp 2, even on one thread. Since modern engines do understand something about trading when ahead, I'm not fully sure why this should be so, but that's the case. Humans know to "keep it simple" against a stronger opponent giving a handicap, but the engines have at least some sense of this, so it's a bit puzzling to me.
[/quote]

Thanks that is what I wanted to know what you stated here "Against humans, Dragon giving knight odds, which means playing White, recent dragon MCTS on 32 threads was clearly too strong for GM Lenderman at 2' + 1" (he needed at least 4' + 1" to score decently as Black), might be close with Nakamura"

[lkaufman]

Here we clearly see GM Nakamura beating all the chess.com bots

Those are not all the chess.com bots, they seem to stop at elo 2100. On the show I watched, he played bots in the 2500 to 2750 elo range and generally had a tough time. He scored somewhere around even with the ones in the 2500s, but when he had to play "himself" (the Hikaru bot), he lost repeatedly, I don't remember if he finally won a game or not. These bots would be several hundred elo lower on CCRL blitz scale.

That is true, GM Nakamura had to tried more than 5 times in order to beat GM Maurice Bot rated around 2500. That is very funny to say the least, I know that a bot simulating to play like yourself it is not even close since human with pressure blunder and GM Nakamura will never beat a bot of GM Nakamura strength

Note: Based on your estimate Dragon playing at TC 2'+1" with Knight Odds is probably rated around 2350 or higher ? I can not wait until dragon beat GM Nakamura at Blitz 2'+1" with knight Odds

I don't know what list "2350" refers to. In engine vs engine play at 2' + 1", a recent Dragon dev. version on one thread performed at 2903 CCRL blitz giving knight odds to Wasp 2 on one thread, and at 2961 with both using four threads (several hundred games each). Against humans, Dragon giving knight odds, which means playing White, recent dragon MCTS on 32 threads was clearly too strong for GM Lenderman at 2' + 1" (he needed at least 4' + 1" to score decently as Black), might be close with Nakamura, who is usually upper 3100s chess.com blitz, 2900 FIDE blitz, and perhaps something like 2300 CCRL blitz scale. Obviously there is a huge difference between giving knight odds to a human and giving knight odds to an engine, since I'm quite sure that Nakamura would get crushed playing 2' + 1" chess on even terms with Wasp 2, even on one thread. Since modern engines do understand something about trading when ahead, I'm not fully sure why this should be so, but that's the case. Humans know to "keep it simple" against a stronger opponent giving a handicap, but the engines have at least some sense of this, so it's a bit puzzling to me.
[/quote]

Thanks that is what I wanted to know what you stated here "Against humans, Dragon giving knight odds, which means playing White, recent dragon MCTS on 32 threads was clearly too strong for GM Lenderman at 2' + 1" (he needed at least 4' + 1" to score decently as Black), might be close with Nakamura"
[/quote]

Some new data. By trying many values, I determined that the latest Lc0 set to a max of 1.5 nodes per second (cpu version, one thread, largest size net) is about an even match for latest Komodo Dragon giving knight odds at 2' + 1" (single thread). Then I determined an approximate CCRL blitz rating for that setting of Lc0; it came out 2450. So although Dragon can give knight odds with even results in blitz to a 2900 CCRL rated standard (alpha-beta) engine, it can only do so to a 2450 pure NN engine. Quite a dramatic difference! Furthermore, the 1.5 nps setting was almost dead even (-7 elo in 500 games) against Komodo 14.1 Skill 23, which is the setting used for the top bots (Nakamura, So, Anand, Giri, Kramnik) on chess.com. But Nakamura could not win any blitz games from the Nakamura bot on the show (he could probably beat it in a Rapid match, but here the topic is blitz), so most likely he would also have very little chance playing blitz with Lc0 set to 1.5 NPS. Yet he can probably score well at knight odds vs. Dragon in 2' + 1" blitz, so it seems that even Lc0 is not a fully satisfactory substitute for a human for knight odds play, though a much better one than normal engines. I think that Nakamura could beat Lc0 in blitz if set to just one node per move (0.1 NPS will do that in blitz), but Lc0 at that setting has no chance vs. Dragon with knight odds. Lc0 does seem to be the best "artificial human" available now for testing dragon giving knight odds.

[Chessqueen]

Here we clearly see GM Nakamura beating all the chess.com bots

Those are not all the chess.com bots, they seem to stop at elo 2100. On the show I watched, he played bots in the 2500 to 2750 elo range and generally had a tough time. He scored somewhere around even with the ones in the 2500s, but when he had to play "himself" (the Hikaru bot), he lost repeatedly, I don't remember if he finally won a game or not. These bots would be several hundred elo lower on CCRL blitz scale.

That is true, GM Nakamura had to tried more than 5 times in order to beat GM Maurice Bot rated around 2500. That is very funny to say the least, I know that a bot simulating to play like yourself it is not even close since human with pressure blunder and GM Nakamura will never beat a bot of GM Nakamura strength

Note: Based on your estimate Dragon playing at TC 2'+1" with Knight Odds is probably rated around 2350 or higher ? I can not wait until dragon beat GM Nakamura at Blitz 2'+1" with knight Odds

I don't know what list "2350" refers to. In engine vs engine play at 2' + 1", a recent Dragon dev. version on one thread performed at 2903 CCRL blitz giving knight odds to Wasp 2 on one thread, and at 2961 with both using four threads (several hundred games each). Against humans, Dragon giving knight odds, which means playing White, recent dragon MCTS on 32 threads was clearly too strong for GM Lenderman at 2' + 1" (he needed at least 4' + 1" to score decently as Black), might be close with Nakamura, who is usually upper 3100s chess.com blitz, 2900 FIDE blitz, and perhaps something like 2300 CCRL blitz scale. Obviously there is a huge difference between giving knight odds to a human and giving knight odds to an engine, since I'm quite sure that Nakamura would get crushed playing 2' + 1" chess on even terms with Wasp 2, even on one thread. Since modern engines do understand something about trading when ahead, I'm not fully sure why this should be so, but that's the case. Humans know to "keep it simple" against a stronger opponent giving a handicap, but the engines have at least some sense of this, so it's a bit puzzling to me.

Thanks that is what I wanted to know what you stated here "Against humans, Dragon giving knight odds, which means playing White, recent dragon MCTS on 32 threads was clearly too strong for GM Lenderman at 2' + 1" (he needed at least 4' + 1" to score decently as Black), might be close with Nakamura"
[/quote]

Some new data. By trying many values, I determined that the latest Lc0 set to a max of 1.5 nodes per second (cpu version, one thread, largest size net) is about an even match for latest Komodo Dragon giving knight odds at 2' + 1" (single thread). Then I determined an approximate CCRL blitz rating for that setting of Lc0; it came out 2450. So although Dragon can give knight odds with even results in blitz to a 2900 CCRL rated standard (alpha-beta) engine, it can only do so to a 2450 pure NN engine. Quite a dramatic difference! Furthermore, the 1.5 nps setting was almost dead even (-7 elo in 500 games) against Komodo 14.1 Skill 23, which is the setting used for the top bots (Nakamura, So, Anand, Giri, Kramnik) on chess.com. But Nakamura could not win any blitz games from the Nakamura bot on the show (he could probably beat it in a Rapid match, but here the topic is blitz), so most likely he would also have very little chance playing blitz with Lc0 set to 1.5 NPS. Yet he can probably score well at knight odds vs. Dragon in 2' + 1" blitz, so it seems that even Lc0 is not a fully satisfactory substitute for a human for knight odds play, though a much better one than normal engines. I think that Nakamura could beat Lc0 in blitz if set to just one node per move (0.1 NPS will do that in blitz), but Lc0 at that setting has no chance vs. Dragon with knight odds. Lc0 does seem to be the best "artificial human" available now for testing dragon giving knight odds.
[/quote]


Next time when you get a GM to play against your Komodo Dragon, do NOT forget to offer them a Nice Coffee Drink, and possibly you can sue Starbuck for using that Name for their coffee
https://www.grocery.com/store/starbucks ... gLQ9_D_BwE