Eval of mainline King's Indian

[EroSennin]

Here is what I would like as white to get.


[pgn]1. d4 Nf6 2. c4 g6 3. Nc3 Bg7 4. e4 d6 5. Nf3 O-O 6. Be2 e5 7. O-O Nc6 8. d5 Ne7 9. Ne1 Nd7 10. Nd3 f5 11. f3 f4 12. Bd2 g5 13. Rc1 Nf6 14. c5 Ng6 15. Nb5 Rf7 16. Ba5 b6 17. cxd6 cxd6 18. Be1 a6 19. Nc3
[/pgn]

Every game Nakamura has won as black after 9.Ne1, computer has shown white to be like plus 2 at some point. Then white made some mistakes and got mated. So anyway I understand king's indian is too difficult for humans to handle. But I still fear from the theoretical viewpoint. I have looked at some lines 30 moves in and computer just keeps preferring white by at least 0.5 points.

[Don]

This opening is notoriously hard to evaluate by computers. The problem is, Black's plan is basically a very gradual but often deadly attack on the kingside. If White parries this he is ok. If White doesn't take effective measures in time, he will lose. This kind of super long-range planning is not something current chess programs do well.

--Jon

You are describing a position that is hard to play but the question is whether it can be evaluated with a more or less correct score - one that sums up the chances well. Surprisingly, the two things are not always the same. A program can often win positions (i.e. play the right moves) without even knowing a position is a win.

In this case, you make the argument that the computer doesn't know that white has a good position but I don't think that is the problem because programs overestimate whites potential. The problem is that they do not understand that black has resources.

I think that you maybe did not interpret that correctly, because you seem to be arguing the same thing here Don, as Jon does. The problem is not with the evaluation of White, because the program can, by measuring mobility, see that White has a lot of extra space. But it is much harder for a program to see if Black could still mobilize his forces quickly once it is clear where the pieces have to go. That is much harder to see in the short term. And what from Jon's description would follow that there is at least part of a King safety problem, which is encouraging if you look at the position at the end of Sune's line

Perhaps. However I would point out that the good program have a fairly well developed sense of king safety and I know that Komodo understand pawn storms and such. These "slow burning" attacks are indeed a challenge for computers though - I realize that.

These type of position do have king safety issues even with the top programs but probably just as importantly I believe that what computers really have a difficult time understanding is when one side looks awkward and cramped and yet is perfectly solid with long term prospects to unwind their pieces later. There are some positions that are considered theoretically equal or with a normal white advantage and yet programs such as Stockfish give them 70 points or so. The french advance and Pirc come to mind - solid black positions where white's advantage is relatively normal and black has counter-attack resources.

[D]r1bq1rk1/pppnn1bp/3p2p1/3Pp3/2P1Pp2/2N2P2/PP2BBPP/R2QNRK1 b - -
This should look familiar for the readers of Lyudmil Tsvetkov's posts, the outline of at least a pawn formation that could lead to King Safety asymmetry: the "sniper wall" rears its ugly head again. Any attacks on Black's shortcastled position would be blocked effectively, but if Black could get some pieces looking in the direction of the long diagonal c8 - h3 the white King is not yet under direct attack, but there is indirect danger I think. I am sure the structure is not hard to program for you Don. And I am even more sure Larry would love to play with it

I can't guarantee the program would actually play better though, if it recognized the structure. In this case it is also about long rang planning if I read Jon's description. And the mobilization of Black's forces.

Eelco

This is something that Larry and I are definitely interested in because we are keenly interested in improving Komodo's understanding of all kinds of difficult chess positions.

[jdart]

I wouldn't worry too much about the eval here. The root problem is that Black's compensation will show up much later in the game. You cannot expect to see that with a static eval or a moderate-depth search.

Another example: my program comes out of the QGA usually with a pretty strong minus score for White because it doesn't think the compensation is worth a pawn. I could fix that but in fact it plays games from these positions reasonably well.

That said I agree with Eelco that a more accurate king safety eval might help. Here is a sample game where Black's attack succeeds:

http://www.chessgames.com/perl/chessgame?gid=1070485

Black applies the classic clamp and then attacks (Kasparov was really fearsome with the KID).

[Dr.Wael Deeb]

I wouldn't worry too much about the eval here. The root problem is that Black's compensation will show up much later in the game. You cannot expect to see that with a static eval or a moderate-depth search.

Another example: my program comes out of the QGA usually with a pretty strong minus score for White because it doesn't think the compensation is worth a pawn. I could fix that but in fact it plays games from these positions reasonably well.

That said I agree with Eelco that a more accurate king safety eval might help. Here is a sample game where Black's attack succeeds:

http://www.chessgames.com/perl/chessgame?gid=1070485

Black applies the classic clamp and then attacks (Kasparov was really fearsome with the KID).

I am a hardcore KID player along with the Old Indian Defence....

I will soon create a special thread where I'll share some of my games against chess engines......
Dr.D

[Kyodai]

It's rather easy to win (or score) in KID vs engines - as this is an opening
that most engines have difficult to "understand".

After a few games it just gets boring to repeat this opening vs engines -
since you get the same kind of advantage most every game.

An exception might be games vs say Houdini and with a special and huge
KID-book. That of course is a challenge - but most of the challenge consists
of playing vs the human knowledge of KID i.e. the human "theory".

Depends on what your aim is. If it's "look look I beat Houfish" - then
repeating KID is a way - but I find this path too boring and much too
limiting concerning the own creativity...

[Don]

I wouldn't worry too much about the eval here. The root problem is that Black's compensation will show up much later in the game. You cannot expect to see that with a static eval or a moderate-depth search.

Another example: my program comes out of the QGA usually with a pretty strong minus score for White because it doesn't think the compensation is worth a pawn. I could fix that but in fact it plays games from these positions reasonably well.

That said I agree with Eelco that a more accurate king safety eval might help. Here is a sample game where Black's attack succeeds:

http://www.chessgames.com/perl/chessgame?gid=1070485

Black applies the classic clamp and then attacks (Kasparov was really fearsome with the KID).

I am a hardcore KID player along with the Old Indian Defence....

I will soon create a special thread where I'll share some of my games against chess engines......
Dr.D

I am particularly interested in seeing those games in order to improve Komodo. Larry and I both understand how badly program in general are in certain types of closed positions and want to do work on that.

[Dr.Wael Deeb]

It's rather easy to win (or score) in KID vs engines - as this is an opening
that most engines have difficult to "understand".

After a few games it just gets boring to repeat this opening vs engines -
since you get the same kind of advantage most every game.

An exception might be games vs say Houdini and with a special and huge
KID-book. That of course is a challenge - but most of the challenge consists
of playing vs the human knowledge of KID i.e. the human "theory".

Depends on what your aim is. If it's "look look I beat Houfish" - then
repeating KID is a way - but I find this path too boring and much too
limiting concerning the own creativity...

In general I agree,but there are hundreds & hundreds of opening lines in the KID that you have to play before you reach the boring cap.....

Besides,not all of the engines in all of the opening lines of the KID play bad or don't understand the positional concept of this opening system....

You don't have the ultimate advantage playing the black pieces.At best you have a solid to slightly better position which is not easy at all to transfer into an easy win

And I am talking about engines which are below the 2700 Elo barrier....

The latest victim I've played was Alfil 8.1.1 Optimised but it didn't go down without a tough fight though
Dr.D

[Dr.Wael Deeb]

I wouldn't worry too much about the eval here. The root problem is that Black's compensation will show up much later in the game. You cannot expect to see that with a static eval or a moderate-depth search.

Another example: my program comes out of the QGA usually with a pretty strong minus score for White because it doesn't think the compensation is worth a pawn. I could fix that but in fact it plays games from these positions reasonably well.

That said I agree with Eelco that a more accurate king safety eval might help. Here is a sample game where Black's attack succeeds:

http://www.chessgames.com/perl/chessgame?gid=1070485

Black applies the classic clamp and then attacks (Kasparov was really fearsome with the KID).

I am a hardcore KID player along with the Old Indian Defence....

I will soon create a special thread where I'll share some of my games against chess engines......
Dr.D

I am particularly interested in seeing those games in order to improve Komodo. Larry and I both understand how badly program in general are in certain types of closed positions and want to do work on that.

I'll create a special thread tomorrow where I will provide games played against chess engines applying the KID and I'll try to highlight the key positions during the games as I saw them OTB and after that....
Dr.D

[carldaman]

28...Nh1! to end the game, superb finale to a great attack!

Thanks for the link,
CL

[Dicaste]

My private engine "Control" prefers "c6" instead of "f4".

Code: Select all

FEN: r1bq1rk1/pppnn1bp/3p2p1/3Ppp2/2P1P3/2NN1P2/PP2B1PP/R1BQ1RK1 b - - 0 11 

Control x64:
  4/17	00:00.109	       3,007	1,002,000	-0.78	fxe4 fxe4 Rxf1+ Qxf1 Nb6 b3
  5/17	00:00.113	       4,091	1,022,000	-0.86	fxe4 fxe4 Rxf1+ Qxf1 Nb6 b3 Nd7 Bg5
  6/17	00:00.117	       5,827	832,000	-0.86	fxe4 fxe4 Rxf1+ Qxf1 Nb6 b3 Nd7 Bg5
  7/18	00:00.122	       9,314	846,000	-0.86	fxe4 fxe4 Rxf1+ Qxf1 Nb6 b3 Nd7 Bg5 a5
  7/18	00:00.128	      13,919	773,000	-0.69	Rf7 Bg5 f4 Qc2 Bf6 Bh6 g5
  8/22	00:00.299	      38,106	129,000	-0.58	Rf7 Bg5 Bf6 Be3 a5 Qc2 f4 Bf2 b6
  9/25	00:00.627	      65,328	104,000	-0.58	Rf7 Bg5 Bf6 Be3 a5 Qc2 f4 Bf2 b6
 10/27	00:00.777	     197,865	255,000	-0.53	Rf7 Bg5 Bf6 Be3 f4 Bf2 b6 Qc2 a5 Nb5
 11/29	00:01.064	     541,737	511,000	-0.59	Rf7 Bg5 Bf6 Be3 f4 Bf2 b6 Qc2 a5 Nb5 Bb7
 11/29	00:01.175	     684,339	584,000	-0.54	a5 Be3 f4 Bf2 b6 Rc1 Ba6 Nb5 Bf6 Qc2 Bb7 Rcd1
 12/31	00:01.517	   1,293,571	854,000	-0.50	a5 Be3 b6 Rc1 Bb7 Qd2 f4 Bf2 Bf6 Nb5 Rc8 b3
 13/31	00:01.911	   2,082,245	1,091,000	-0.52	a5 Be3 b6 Qd2 Bb7 b3 Rc8 Rac1 f4 Bf2 Nf6 Nb5 Qd7 Bh4
 13/32	00:02.276	   2,947,747	1,297,000	-0.49	f4 b3 b6 Bb2 Bb7 Qd2 Nf6 Rac1 c6 Nf2 Rc8 dxc6 Rxc6 Nb5 a6
 14/32	00:02.464	   3,297,953	1,340,000	-0.44	f4 b3 b6 Bb2 Bb7 Qd2 Nf6 Rac1 c6 Nb4 c5 Nd3 a6 Nf2
 15/33	00:03.811	   7,007,430	1,840,000	-0.45	f4 b3 b6 Bb2 Bb7 Qd2 Nf6 Nf2 Qd7 Rac1 Rac8 Rcd1 c6 dxc6 Rxc6 Nb5
 16/34	00:05.442	  14,359,620	2,640,000	-0.43	f4 b3 b6 Bb2 Bb7 Qd2 Nf6 Nf2 Qd7 Rac1 Rac8 Rcd1 a6 b4 a5 bxa5
 17/36	00:08.164	  25,834,273	3,165,000	-0.45	f4 b3 b6 Bb2 Bb7 Qd2 Nf6 Nf2 Qd7 Nh3 Bh6 Bd3 a6 Nf2 Bg7 Rae1 Rac8
 18/36	00:11.586	  41,507,648	3,583,000	-0.45	f4 b3 b6 Bb2 Bb7 Qd2 Nf6 Nf2 Qd7 Nh3 Bh6 Bd3 a6 Nf2 Bg7 Rae1 Rac8
 18/42	00:44.303	 188,552,410	4,256,000	-0.44	c6 Be3 Nf6 Qb3 c5 Qa3 a5 Qb3 Bd7 Qc2 Qb6 Bg5 fxe4 fxe4 Qa6 b3 a4 Qb2 b5
 19/43	00:51.890	 222,372,816	4,285,000	-0.43	c6 Be3 Nf6 Qb3 c5 a3 a6 Qc2 Bd7 b4 b6 Rab1 Rb8 Bg5 h6 Bh4 Qc7 bxc5 bxc5 Qd2
 19/47	01:35.286	 407,188,294	4,273,000	-0.43	c6 Be3 Nf6 Qb3 c5 a3 a6 Qc2 Bd7 b4 b6 Rab1 Rb8 Bg5 h6 Bh4 Qc7 bxc5 bxc5 Qd2
 19/47	02:28.202	 407,188,294	2,747,000	-0.43	c6 Be3 Nf6 Qb3 c5 a3 a6 Qc2 Bd7 b4 b6 Rab1 Rb8 Bg5 h6 Bh4 Qc7 bxc5 bxc5 Qd2
Best move: c6, Value: -0.43, Depth: 19/47, Time: 02:28.223, 407,188,294 Nodes, 2,747 kN/sec.