Komodo Dragon 2 released.

[Vinvin]

Note :Here is a well known position for illustrate:
1rb1qrk1/2b2pp1/p3pBn1/3pP1Pp/1ppP4/2P1QN2/PP3P1P/R2BR1K1 w - - 0 1

Only Lc0 can find it quickly because it was simply in one of his game ! Stockfish see nothing if you don't help , same thing for Dragon....
I found only one derivative which find the good move at low depth ( 27 or max 30 anyway ) but i use always Muti PV ???

Note that now, one more engine finds Nh4 : SugaR AI ICCF 1.90
3m17s, 1 core on an old i7 :

15 Nh4; id "Hard-Talkc
Searching move: Nf3-h4
Best move (SugaR AI ICCF 1.90 sse41-popcnt): Nf3-h4
Identical moves! Found in: 03:17
Found 511 tablebases
NNUE evaluation using nn-62ef826d1a6d.nnue enabled
...
26/46 02:21 160 340 937 1 135 929 +0,75 1.h4 Kh7 2.Nh2 Rh8 3.Bxh5 Kg8 4.Bxg6 fxg6 5.Ng4 Qf7 6.Kg2 Bd7 7.Rh1 bxc3 8.bxc3 Ba4 9.Qd2 Kf8 10.Rae1 Ke8 11.Re2 Kd7 12.Qc1 Bc6 13.Re3 gxf6 14.exf6 Bf4 15.Qc2 Kc8 16.Ne5 Bxe5 17.Rxe5 Rb7 18.Qe2 Kd7
27/46+ 03:17 224 232 786 1 132 500 +0,89 1.Nh4
27/46+ 03:47 257 801 283 1 133 511 +0,97 1.Nh4
27/52+ 09:03 615 869 794 1 133 386 +1,85 1.Nh4
27/52 10:41 726 833 890 1 133 480 +1,50 1.Nh4 Kh7 2.Bxh5 Rh8 3.Qg3 Kg8 4.Qg4 Bd7 5.Re3 Ba4 6.b3 Bd7 7.cxb4 Rh7 8.Rh3 Rxb4 9.Rb1 Ba5 10.Rc1 Rb8 11.Rc2 Bb6 12.bxc4 Ba4 13.Bxg6 fxg6 14.Rcc3 gxf6 15.gxf6
28/37- 12:26 845 049 974 1 131 746 +1,46 1.Nh4 Kh7
28/49+ 12:48 870 216 964 1 131 756 +1,50 1.Nh4
28/49+ 13:16 902 080 543 1 132 062 +1,65 1.Nh4
28/49 16:46 1 138 146 459 1 131 290 +1,70 1.Nh4 Kh7 2.Bxh5 Rh8 3.Qg3 Kg8 4.Qg4 Bd7 5.Re3 Ba4 6.Bxg6 fxg6 7.Rc1 Bd7 8.Ng2 a5 9.h4 a4 10.a3 bxc3 11.bxc3 Rb3 12.Nf4 Rh7 13.Rh3 Rh5 14.Rd1 Rxa3 15.Nxh5 gxh5 16.Qg2
29/38- 16:55 1 148 666 882 1 131 246 +1,86 1.Nh4 Kh7

[jr66]

Note :Here is a well known position for illustrate:
1rb1qrk1/2b2pp1/p3pBn1/3pP1Pp/1ppP4/2P1QN2/PP3P1P/R2BR1K1 w - - 0 1

Only Lc0 can find it quickly because it was simply in one of his game ! Stockfish see nothing if you don't help , same thing for Dragon....
I found only one derivative which find the good move at low depth ( 27 or max 30 anyway ) but i use always Muti PV ???

Note that now, one more engine finds Nh4 : SugaR AI ICCF 1.90
3m17s, 1 core on an old i7 :

15 Nh4; id "Hard-Talkc
Searching move: Nf3-h4
Best move (SugaR AI ICCF 1.90 sse41-popcnt): Nf3-h4
Identical moves! Found in: 03:17
Found 511 tablebases
NNUE evaluation using nn-62ef826d1a6d.nnue enabled
...
26/46 02:21 160 340 937 1 135 929 +0,75 1.h4 Kh7 2.Nh2 Rh8 3.Bxh5 Kg8 4.Bxg6 fxg6 5.Ng4 Qf7 6.Kg2 Bd7 7.Rh1 bxc3 8.bxc3 Ba4 9.Qd2 Kf8 10.Rae1 Ke8 11.Re2 Kd7 12.Qc1 Bc6 13.Re3 gxf6 14.exf6 Bf4 15.Qc2 Kc8 16.Ne5 Bxe5 17.Rxe5 Rb7 18.Qe2 Kd7
27/46+ 03:17 224 232 786 1 132 500 +0,89 1.Nh4
27/46+ 03:47 257 801 283 1 133 511 +0,97 1.Nh4
27/52+ 09:03 615 869 794 1 133 386 +1,85 1.Nh4
27/52 10:41 726 833 890 1 133 480 +1,50 1.Nh4 Kh7 2.Bxh5 Rh8 3.Qg3 Kg8 4.Qg4 Bd7 5.Re3 Ba4 6.b3 Bd7 7.cxb4 Rh7 8.Rh3 Rxb4 9.Rb1 Ba5 10.Rc1 Rb8 11.Rc2 Bb6 12.bxc4 Ba4 13.Bxg6 fxg6 14.Rcc3 gxf6 15.gxf6
28/37- 12:26 845 049 974 1 131 746 +1,46 1.Nh4 Kh7
28/49+ 12:48 870 216 964 1 131 756 +1,50 1.Nh4
28/49+ 13:16 902 080 543 1 132 062 +1,65 1.Nh4
28/49 16:46 1 138 146 459 1 131 290 +1,70 1.Nh4 Kh7 2.Bxh5 Rh8 3.Qg3 Kg8 4.Qg4 Bd7 5.Re3 Ba4 6.Bxg6 fxg6 7.Rc1 Bd7 8.Ng2 a5 9.h4 a4 10.a3 bxc3 11.bxc3 Rb3 12.Nf4 Rh7 13.Rh3 Rh5 14.Rd1 Rxa3 15.Nxh5 gxh5 16.Qg2
29/38- 16:55 1 148 666 882 1 131 246 +1,86 1.Nh4 Kh7

Many thanks Vinvin,
It was Crystal 3.1 on my side ( i don't know where to find this Sugar ICCF ? )...
Ha, ok sorry : https://github.com/Zerbinati/SugaR-AI-ICCF/releases
By the way, if it is an exception, i am not sure we don't have others exceptions in the 114 Vinvin's positions where "normal" engines are not so good it seem ?
Regards

[CMCanavessi]

...

Larry, when you go to the Komodo website, the link to the free version (Komodo 12?) points to the website again, no download.

[lkaufman]

...

Larry, when you go to the Komodo website, the link to the free version (Komodo 12?) points to the website again, no download.

If you click on "dragon sale" and then "komodo 12 free" it works, but I agree, we should fix the website.

[Chessqueen]

...

Larry, when you go to the Komodo website, the link to the free version (Komodo 12?) points to the website again, no download.

If you click on "dragon sale" and then "komodo 12 free" it works, but I agree, we should fix the website.

By now Komodo 13.3 should be offered for Free or for $10.00
This could be a great dead ==> forum3/viewtopic.php?f=2&t=77649&p=898387#p898387

[Madeleine Birchfield]

Larry Kaufman, have the Komodo team tried out Monte-Carlo Graph Search yet? This paper claims that it is an improvement over Monte-Carlo Tree Search:

https://arxiv.org/abs/2012.11045

[lkaufman]

Larry Kaufman, have the Komodo team tried out Monte-Carlo Graph Search yet? This paper claims that it is an improvement over Monte-Carlo Tree Search:

https://arxiv.org/abs/2012.11045

No, thanks for the idea. Our MCTS doesn't have much of a memory problem, so it might not be very helpful, but I don't know of course.

[KLc]

Here's a rather simple tactical position Dragon 2 in MCTS mode doesn't understand and I wonder what the problem is:

[d]8/pp3p2/5p2/5p1P/1P1bk3/P5B1/2K5/8 w - - 0 1

It wants to play 1.h6. But this move is bad because the first two pawns of the triple pawn group can move forward, thus opening the diagonal for the black bishop which then covers the queening square of the h-pawn, e.g. 1.h6 f4 2.Bf2 Be5 3.Bxa7 b5 4.h7 f5.

But white can easily promote the h-pawn by first sacrificing the bishop: 1.Bf4! Kxf4 2.h6 Ke3 3.h7 b6 4.h8=Q.

What's the problem here?

My settings are:

Code: Select all

Threads=2
Hash=1024
Use MCTS=true
MCTS Hash=1024
Ponder=false
OwnBook=false
Use Syzygy=false

[Chessqueen]

Here's a rather simple tactical position Dragon 2 in MCTS mode doesn't understand and I wonder what the problem is:

[d]8/pp3p2/5p2/5p1P/1P1bk3/P5B1/2K5/8 w - - 0 1

It wants to play 1.h6. But this move is bad because the first two pawns of the triple pawn group can move forward, thus opening the diagonal for the black bishop which then covers the queening square of the h-pawn, e.g. 1.h6 f4 2.Bf2 Be5 3.Bxa7 b5 4.h7 f5.

But white can easily promote the h-pawn by first sacrificing the bishop: 1.Bf4! Kxf4 2.h6 Ke3 3.h7 b6 4.h8=Q.

What's the problem here?

My settings are:

Code: Select all

Threads=2
Hash=1024
Use MCTS=true
MCTS Hash=1024
Ponder=false
OwnBook=false
Use Syzygy=false

Kurt you seems to be a strong player of at least Master strength, based on your analysis. Can you download Honey latest version and set it to play at your level, under Arena you can click Engine setting and select your level ( GM, IM Expert ect) . It would be interesting to see a game to determine if the Levels are set up correctly to play against Human. To download it ==>https://github.com/MichaelB7/Stockfish/releases/tag/14

[lkaufman]

Here's a rather simple tactical position Dragon 2 in MCTS mode doesn't understand and I wonder what the problem is:

[d]8/pp3p2/5p2/5p1P/1P1bk3/P5B1/2K5/8 w - - 0 1

It wants to play 1.h6. But this move is bad because the first two pawns of the triple pawn group can move forward, thus opening the diagonal for the black bishop which then covers the queening square of the h-pawn, e.g. 1.h6 f4 2.Bf2 Be5 3.Bxa7 b5 4.h7 f5.

But white can easily promote the h-pawn by first sacrificing the bishop: 1.Bf4! Kxf4 2.h6 Ke3 3.h7 b6 4.h8=Q.

What's the problem here?

My settings are:

Code: Select all

Threads=2
Hash=1024
Use MCTS=true
MCTS Hash=1024
Ponder=false
OwnBook=false
Use Syzygy=false

The problem is that the MCTS version thinks that the line played, winning the a7 pawn, is pretty good, ultimately winning a bishop but not the game. Also the correct line, sacrificing a bishop for nothing to queen a pawn several moves later, is very hard to see without a deep search, due to pruning. MCTS is a bit weak in seeing sacrifices that don't pay off quickly like this one. It is better at playing against an opponent who is not like itself, especially a human, because it doesn't assume that the opponent will play what it would play. That's the trade-off.