Komodo Dragon 3 Released by KomodoChess

[Fritz 0]

Dragon 3 still blunders inexplicably, similar to Dragon 2.6, even at pretty high Elo settings. Today I played a slow (90+30) game against 2225 Elo, so that I got an about 2000 Elo oponent for that time control. Well, its play was bizarre, to say the least. On the other hand, my play was worth a psyhiatric expertise, but it's not about me now. I will post the game tomorrow and explain my observations. For now I will say that Komodo level 20 (7 ply), which is about that strength in a direct encounter, would never play in such a unnatural way. It makes mistakes too, but not such trivial ones.

[lkaufman]

Dragon 3 still blunders inexplicably, similar to Dragon 2.6, even at pretty high Elo settings. Today I played a slow (90+30) game against 2225 Elo, so that I got an about 2000 Elo oponent for that time control. Well, its play was bizarre, to say the least. On the other hand, my play was worth a psyhiatric expertise, but it's not about me now. I will post the game tomorrow and explain my observations. For now I will say that Komodo level 20 (7 ply), which is about that strength in a direct encounter, would never play in such a unnatural way. It makes mistakes too, but not such trivial ones.

I'll have to look at the game, but note that even Magnus Carlsen just blundered a piece to a simple rook fork in a 15' + 10" game, the time limit Dragon elos are calibrated for, so a 2225 setting should be expected to make some pretty bad blunders on occasion, though it should never just overlook a direct winning capture at that level (although humans of that level do so sometimes in Rapid).

[Fritz 0]

[pgn][Event "Blitz 90m+30s"]
[Date "2022.05.21"]
[Round "?"]
[White "Dragon 3 by Komodo Chess 64-b"]
[Black "Branislav Djosic"]
[Result "1-0"]
[ECO "B07"]
[TimeControl "5400+30"]

{ 128MB, Fritz 16.ctg } 1.d4 d6 2.e4 Nf6 3.Nc3 e5 4.Nge2 Be7 5.f3 O-O 6.Be3 c6 7.a4 b6 8.g4 Ne8 9.d5 Bg5 10.Bf2 Bh4 11.Bxh4 Qxh4+ 12.Ng3 c5 13.Qd2 a6 14.a5 b5 15.Nd1 Nd7 16.Ne3 Ndf6 17.Ng2 Nxe4 18.Nxh4 Nxd2 19.Kxd2 Nf6 20.c4 b4 21.Nhf5 Rd8 22.g5 Ne8 23.Bh3 Kf8 24.Raf1 g6 25.Rf2 gxf5 26.f4 exf4 27.Rxf4 Ng7 28.Rh4 Kg8 29.Rf1 Ra7 30.Rf3 Re7 31.Rh6 Re5 32.Rf4 Kf8 33.Rxh7 Rde8 34.Bf1 Bd7 35.Bd3 Bc8 36.Bc2 Bd7 37.h3 Kg8 38.Rfh4 Kf8 39.b3 Bc8 40.Rf4 Kg8 41.Rfh4 Kf8 42.Bd3 Bd7 43.R4h6 f4 44.Ne4 Nf5 45.Nf6 Rd8 46.Rh8+ Ke7 47.Ng8+ Kf8 48.R6h7 f3 49.Nh6+ Ke7 50.Rxf7# 1-0

[/pgn]

Well, that's the game. First antipositional Dragon's behaviour was to let me play 13. ... a6 and then 14. ... b5. Why not 13. a5 instead of 13. Qd2? It is elementary in such positions. Before that, my 12. ... c5 was probably dubious, but I didn't like the d5 square to be available to the White pieces while I didn't have a Knight on f6.

My 16. ... Ndf6?? is for the urgent hospitalization, no other comment required. Which is the worst, I saw the threat Nd1-Ne3-Ng2 a few moves before, but then forgot about it. After that, I practically gave up, not even considering 17. ... Qh3 and walking into a discovered attack. But...

24. Raf1???. C'mon people, which 2225 Elo player would blunder like that in a 15+10 game?! Or, if you want, which 2000 Elo player would make such an one move nonsense in 90+30? After I telephoned that I intended to play 24. ... g6 by playing 23. ... Kf8? Totally unnatural and unrealistic, if you ask me.

Unfortunately, Black position was too bad and probably lost even after that piece gift. There was some solid techique of Dragon to convert it into the victory, I admit that, but even my late grandmother would win this as White, considering indifensible h7 Pawn and Black practically not being able to move anything.

Maybe I'm wrong, but from now on I consider Komodo 13.3 level 20 a much better simulation of a 2000 Elo player in slow chess. NNUE engines are surely stronger at full strength, but crippled ones play nonsense like this too often. Perhaps NNUE and limited nodes/search depth don't go well together, I don't know, but that's what I see.

[lkaufman]

[pgn][Event "Blitz 90m+30s"]
[Date "2022.05.21"]
[Round "?"]
[White "Dragon 3 by Komodo Chess 64-b"]
[Black "Branislav Djosic"]
[Result "1-0"]
[ECO "B07"]
[TimeControl "5400+30"]

{ 128MB, Fritz 16.ctg } 1.d4 d6 2.e4 Nf6 3.Nc3 e5 4.Nge2 Be7 5.f3 O-O 6.Be3 c6 7.a4 b6 8.g4 Ne8 9.d5 Bg5 10.Bf2 Bh4 11.Bxh4 Qxh4+ 12.Ng3 c5 13.Qd2 a6 14.a5 b5 15.Nd1 Nd7 16.Ne3 Ndf6 17.Ng2 Nxe4 18.Nxh4 Nxd2 19.Kxd2 Nf6 20.c4 b4 21.Nhf5 Rd8 22.g5 Ne8 23.Bh3 Kf8 24.Raf1 g6 25.Rf2 gxf5 26.f4 exf4 27.Rxf4 Ng7 28.Rh4 Kg8 29.Rf1 Ra7 30.Rf3 Re7 31.Rh6 Re5 32.Rf4 Kf8 33.Rxh7 Rde8 34.Bf1 Bd7 35.Bd3 Bc8 36.Bc2 Bd7 37.h3 Kg8 38.Rfh4 Kf8 39.b3 Bc8 40.Rf4 Kg8 41.Rfh4 Kf8 42.Bd3 Bd7 43.R4h6 f4 44.Ne4 Nf5 45.Nf6 Rd8 46.Rh8+ Ke7 47.Ng8+ Kf8 48.R6h7 f3 49.Nh6+ Ke7 50.Rxf7# 1-0

[/pgn]

Well, that's the game. First antipositional Dragon's behaviour was to let me play 13. ... a6 and then 14. ... b5. Why not 13. a5 instead of 13. Qd2? It is elementary in such positions. Before that, my 12. ... c5 was probably dubious, but I didn't like the d5 square to be available to the White pieces while I didn't have a Knight on f6.

My 16. ... Ndf6?? is for the urgent hospitalization, no other comment required. Which is the worst, I saw the threat Nd1-Ne3-Ng2 a few moves before, but then forgot about it. After that, I practically gave up, not even considering 17. ... Qh3 and walking into a discovered attack. But...

24. Raf1???. C'mon people, which 2225 Elo player would blunder like that in a 15+10 game?! Or, if you want, which 2000 Elo player would make such an one move nonsense in 90+30? After I telephoned that I intended to play 24. ... g6 by playing 23. ... Kf8? Totally unnatural and unrealistic, if you ask me.

Unfortunately, Black position was too bad and probably lost even after that piece gift. There was some solid techique of Dragon to convert it into the victory, I admit that, but even my late grandmother would win this as White, considering indifensible h7 Pawn and Black practically not being able to move anything.

Maybe I'm wrong, but from now on I consider Komodo 13.3 level 20 a much better simulation of a 2000 Elo player in slow chess. NNUE engines are surely stronger at full strength, but crippled ones play nonsense like this too often. Perhaps NNUE and limited nodes/search depth don't go well together, I don't know, but that's what I see.

Is there any chance you had the "autoskill" option active? That would explain the 24.Raf1?? blunder; autoskill deliberately lowers the level of play significanlty when the engine is winning to give you a chance, it assumes that it must have been playing at too high a level for you. I would expect Dragon 3 to make a blunder like 24.Raf1?? with Elo set to maybe 1800 to 1900, but not 2225. It is a 3 ply tactic, and normally Dragon should complete 3 plies at 2225 setting. As for the poor move a5, I'll have to investigate.

[Fritz 0]

[pgn][Event "Blitz 90m+30s"]
[Date "2022.05.21"]
[Round "?"]
[White "Dragon 3 by Komodo Chess 64-b"]
[Black "Branislav Djosic"]
[Result "1-0"]
[ECO "B07"]
[TimeControl "5400+30"]

{ 128MB, Fritz 16.ctg } 1.d4 d6 2.e4 Nf6 3.Nc3 e5 4.Nge2 Be7 5.f3 O-O 6.Be3 c6 7.a4 b6 8.g4 Ne8 9.d5 Bg5 10.Bf2 Bh4 11.Bxh4 Qxh4+ 12.Ng3 c5 13.Qd2 a6 14.a5 b5 15.Nd1 Nd7 16.Ne3 Ndf6 17.Ng2 Nxe4 18.Nxh4 Nxd2 19.Kxd2 Nf6 20.c4 b4 21.Nhf5 Rd8 22.g5 Ne8 23.Bh3 Kf8 24.Raf1 g6 25.Rf2 gxf5 26.f4 exf4 27.Rxf4 Ng7 28.Rh4 Kg8 29.Rf1 Ra7 30.Rf3 Re7 31.Rh6 Re5 32.Rf4 Kf8 33.Rxh7 Rde8 34.Bf1 Bd7 35.Bd3 Bc8 36.Bc2 Bd7 37.h3 Kg8 38.Rfh4 Kf8 39.b3 Bc8 40.Rf4 Kg8 41.Rfh4 Kf8 42.Bd3 Bd7 43.R4h6 f4 44.Ne4 Nf5 45.Nf6 Rd8 46.Rh8+ Ke7 47.Ng8+ Kf8 48.R6h7 f3 49.Nh6+ Ke7 50.Rxf7# 1-0

[/pgn]

Well, that's the game. First antipositional Dragon's behaviour was to let me play 13. ... a6 and then 14. ... b5. Why not 13. a5 instead of 13. Qd2? It is elementary in such positions. Before that, my 12. ... c5 was probably dubious, but I didn't like the d5 square to be available to the White pieces while I didn't have a Knight on f6.

My 16. ... Ndf6?? is for the urgent hospitalization, no other comment required. Which is the worst, I saw the threat Nd1-Ne3-Ng2 a few moves before, but then forgot about it. After that, I practically gave up, not even considering 17. ... Qh3 and walking into a discovered attack. But...

24. Raf1???. C'mon people, which 2225 Elo player would blunder like that in a 15+10 game?! Or, if you want, which 2000 Elo player would make such an one move nonsense in 90+30? After I telephoned that I intended to play 24. ... g6 by playing 23. ... Kf8? Totally unnatural and unrealistic, if you ask me.

Unfortunately, Black position was too bad and probably lost even after that piece gift. There was some solid techique of Dragon to convert it into the victory, I admit that, but even my late grandmother would win this as White, considering indifensible h7 Pawn and Black practically not being able to move anything.

Maybe I'm wrong, but from now on I consider Komodo 13.3 level 20 a much better simulation of a 2000 Elo player in slow chess. NNUE engines are surely stronger at full strength, but crippled ones play nonsense like this too often. Perhaps NNUE and limited nodes/search depth don't go well together, I don't know, but that's what I see.

Is there any chance you had the "autoskill" option active? That would explain the 24.Raf1?? blunder; autoskill deliberately lowers the level of play significanlty when the engine is winning to give you a chance, it assumes that it must have been playing at too high a level for you. I would expect Dragon 3 to make a blunder like 24.Raf1?? with Elo set to maybe 1800 to 1900, but not 2225. It is a 3 ply tactic, and normally Dragon should complete 3 plies at 2225 setting. As for the poor move a5, I'll have to investigate.

No, it was not active, for sure.

[Fritz 0]

One more question - Larry, do you think that switching to regular eval while maintaining the same Elo setting would eliminate such blunders?

[lkaufman]

One more question - Larry, do you think that switching to regular eval while maintaining the same Elo setting would eliminate such blunders?

Yes, it would dramatically reduce such tactical blunders, while increasing the number of positional errors. That is the tradeoff. Personally I think that using NNUE with the lower Elo values is more "human" in that strong human players play pretty well positionally even with short time but do make tactical errors. Maybe the engine's tactical errors aren't exactly of the same type as a similarly rated human, but overall I think the simulation is better.

[lkaufman]

[pgn][Event "Blitz 90m+30s"]
[Date "2022.05.21"]
[Round "?"]
[White "Dragon 3 by Komodo Chess 64-b"]
[Black "Branislav Djosic"]
[Result "1-0"]
[ECO "B07"]
[TimeControl "5400+30"]

{ 128MB, Fritz 16.ctg } 1.d4 d6 2.e4 Nf6 3.Nc3 e5 4.Nge2 Be7 5.f3 O-O 6.Be3 c6 7.a4 b6 8.g4 Ne8 9.d5 Bg5 10.Bf2 Bh4 11.Bxh4 Qxh4+ 12.Ng3 c5 13.Qd2 a6 14.a5 b5 15.Nd1 Nd7 16.Ne3 Ndf6 17.Ng2 Nxe4 18.Nxh4 Nxd2 19.Kxd2 Nf6 20.c4 b4 21.Nhf5 Rd8 22.g5 Ne8 23.Bh3 Kf8 24.Raf1 g6 25.Rf2 gxf5 26.f4 exf4 27.Rxf4 Ng7 28.Rh4 Kg8 29.Rf1 Ra7 30.Rf3 Re7 31.Rh6 Re5 32.Rf4 Kf8 33.Rxh7 Rde8 34.Bf1 Bd7 35.Bd3 Bc8 36.Bc2 Bd7 37.h3 Kg8 38.Rfh4 Kf8 39.b3 Bc8 40.Rf4 Kg8 41.Rfh4 Kf8 42.Bd3 Bd7 43.R4h6 f4 44.Ne4 Nf5 45.Nf6 Rd8 46.Rh8+ Ke7 47.Ng8+ Kf8 48.R6h7 f3 49.Nh6+ Ke7 50.Rxf7# 1-0

[/pgn]

Well, that's the game. First antipositional Dragon's behaviour was to let me play 13. ... a6 and then 14. ... b5. Why not 13. a5 instead of 13. Qd2? It is elementary in such positions. Before that, my 12. ... c5 was probably dubious, but I didn't like the d5 square to be available to the White pieces while I didn't have a Knight on f6.

My 16. ... Ndf6?? is for the urgent hospitalization, no other comment required. Which is the worst, I saw the threat Nd1-Ne3-Ng2 a few moves before, but then forgot about it. After that, I practically gave up, not even considering 17. ... Qh3 and walking into a discovered attack. But...

24. Raf1???. C'mon people, which 2225 Elo player would blunder like that in a 15+10 game?! Or, if you want, which 2000 Elo player would make such an one move nonsense in 90+30? After I telephoned that I intended to play 24. ... g6 by playing 23. ... Kf8? Totally unnatural and unrealistic, if you ask me.

Unfortunately, Black position was too bad and probably lost even after that piece gift. There was some solid techique of Dragon to convert it into the victory, I admit that, but even my late grandmother would win this as White, considering indifensible h7 Pawn and Black practically not being able to move anything.

Maybe I'm wrong, but from now on I consider Komodo 13.3 level 20 a much better simulation of a 2000 Elo player in slow chess. NNUE engines are surely stronger at full strength, but crippled ones play nonsense like this too often. Perhaps NNUE and limited nodes/search depth don't go well together, I don't know, but that's what I see.

I checked out the game a bit more. 13.a5 is the "human" move, but Komodo is not convinced that it is actually better than Qd2 even at full strength. But playing 14.a5? is clearly bad, no question. However I can't get Dragon to replicate that bad move with any Elo setting, although with low values like 1000 it will eventually try all sorts of silly moves like this. I can't imagine or explain how it could play it at 2225 setting. As for the later piece blunder on move 24, that seems to be due to Dragon thinking the position is so winning that it can just let you win the knight and still it will win easily; you might call that "trolling"! It does seem to play rather randomly sometimes in totally winning positions, as if trying to let you think you might have a chance. That's related more to the NN training than to the level setting.

[Fritz 0]

[pgn][Event "Blitz 90m+30s"]
[Date "2022.05.21"]
[Round "?"]
[White "Dragon 3 by Komodo Chess 64-b"]
[Black "Branislav Djosic"]
[Result "1-0"]
[ECO "B07"]
[TimeControl "5400+30"]

{ 128MB, Fritz 16.ctg } 1.d4 d6 2.e4 Nf6 3.Nc3 e5 4.Nge2 Be7 5.f3 O-O 6.Be3 c6 7.a4 b6 8.g4 Ne8 9.d5 Bg5 10.Bf2 Bh4 11.Bxh4 Qxh4+ 12.Ng3 c5 13.Qd2 a6 14.a5 b5 15.Nd1 Nd7 16.Ne3 Ndf6 17.Ng2 Nxe4 18.Nxh4 Nxd2 19.Kxd2 Nf6 20.c4 b4 21.Nhf5 Rd8 22.g5 Ne8 23.Bh3 Kf8 24.Raf1 g6 25.Rf2 gxf5 26.f4 exf4 27.Rxf4 Ng7 28.Rh4 Kg8 29.Rf1 Ra7 30.Rf3 Re7 31.Rh6 Re5 32.Rf4 Kf8 33.Rxh7 Rde8 34.Bf1 Bd7 35.Bd3 Bc8 36.Bc2 Bd7 37.h3 Kg8 38.Rfh4 Kf8 39.b3 Bc8 40.Rf4 Kg8 41.Rfh4 Kf8 42.Bd3 Bd7 43.R4h6 f4 44.Ne4 Nf5 45.Nf6 Rd8 46.Rh8+ Ke7 47.Ng8+ Kf8 48.R6h7 f3 49.Nh6+ Ke7 50.Rxf7# 1-0

[/pgn]

Well, that's the game. First antipositional Dragon's behaviour was to let me play 13. ... a6 and then 14. ... b5. Why not 13. a5 instead of 13. Qd2? It is elementary in such positions. Before that, my 12. ... c5 was probably dubious, but I didn't like the d5 square to be available to the White pieces while I didn't have a Knight on f6.

My 16. ... Ndf6?? is for the urgent hospitalization, no other comment required. Which is the worst, I saw the threat Nd1-Ne3-Ng2 a few moves before, but then forgot about it. After that, I practically gave up, not even considering 17. ... Qh3 and walking into a discovered attack. But...

24. Raf1???. C'mon people, which 2225 Elo player would blunder like that in a 15+10 game?! Or, if you want, which 2000 Elo player would make such an one move nonsense in 90+30? After I telephoned that I intended to play 24. ... g6 by playing 23. ... Kf8? Totally unnatural and unrealistic, if you ask me.

Unfortunately, Black position was too bad and probably lost even after that piece gift. There was some solid techique of Dragon to convert it into the victory, I admit that, but even my late grandmother would win this as White, considering indifensible h7 Pawn and Black practically not being able to move anything.

Maybe I'm wrong, but from now on I consider Komodo 13.3 level 20 a much better simulation of a 2000 Elo player in slow chess. NNUE engines are surely stronger at full strength, but crippled ones play nonsense like this too often. Perhaps NNUE and limited nodes/search depth don't go well together, I don't know, but that's what I see.

I checked out the game a bit more. 13.a5 is the "human" move, but Komodo is not convinced that it is actually better than Qd2 even at full strength. But playing 14.a5? is clearly bad, no question. However I can't get Dragon to replicate that bad move with any Elo setting, although with low values like 1000 it will eventually try all sorts of silly moves like this. I can't imagine or explain how it could play it at 2225 setting. As for the later piece blunder on move 24, that seems to be due to Dragon thinking the position is so winning that it can just let you win the knight and still it will win easily; you might call that "trolling"! It does seem to play rather randomly sometimes in totally winning positions, as if trying to let you think you might have a chance. That's related more to the NN training than to the level setting.

Dragon 3 at regular eval Elo 2225 also preferes 13. Qd2, but plays 14. 0-0-0 instead of a5. Naturally, on move 24 doesn't hang a piece, but plays Bg4 instead.

Komodo 13.3 level 20, which is just 5-10 points ahead of Dragon 3 Elo 2225 (with NNUE) in a direct match, plays 13. Qd2 or 13. Kf2?! (very suspicious and odd in my opinion, but I remember that some versions of Fritz at a comparable search depth also played this move in similar positions), 14. 0-0-0 (after 13. Qd2 of course) and 24. Bg4.

But there is a surprising thing - Komodo 13.3 in the infinite analysis seems to prefer 13. Qd2 all the way to 19 ply, then switches to a5 at 20 ply, and then suggests Kf2 at 21 ply!

All of this applies to one thread.

[KLc]

Elo levels have been much more carefully tuned and tested, especially it has been verified that Elo 2600 is a good match for Grandmaster Alex Lenderman (who had that FIDE rating) in 15' + 10" Rapid games. The play has been made more human-like at these Elo settings; at the higher levels there should now be few really obvious blunders. UCI Elo functioning has been restored to the method used in 2.6 rather than 2.6.1 by removing the underscore from the name, as the underscore caused problems.

I don't understand how the underscore can cause a problem. The correct option names have an underscore, e.g. for Stockfish:

Code: Select all

option name UCI_LimitStrength type check default false
option name UCI_Elo type spin default 1350 min 1350 max 2850

Because the underscore is removed the Elo option is not recognized as such in GUIs like HIARCS Chess Explorer. This is not what I expected.