TalkChess.com

A few days ago I was talking to GM Daniel Alsina and told me something that is well known about the inability of the engines to understand some structures in some opening or defenses, like the typical white pawns on f3, e4, d5, and black d6, e5, f4 in the King indian. He told that himself and other GM's just switch off the engine or ignore it on this type of positions as they don't understand the long term plans. Also told about some positions in the French defense, but as in those white was playing only for two results (win or draw), that was not that terrible as in the KID, when you will lose by attack.

When the plan requires a lot of quiet moves and creation of good coordination for only later crashing trhough the position, the current engines are not able to do this as they are optimized for tactics, which of course reign in most of the positions but not all.

Anyway he insisted and I told him that I will investigate and try to do something about it, even if I know and I told that this possibly requires some new concepts to appear, as we are talking that with current engine techniques maybe they will require at least 10 or 15 plies more to simulate that they understand this type of positions, who knows.

Maybe this can be aborded by tuning some parameters specifically for this type of structures. Or instead of tuning parameters, something can be done adding a helper function like the ones we tend to do for the endgames.

In some time I will try to investigate, but I leave this written here for if the possible discussion gives something.

Probably because of self-testing to find improvements.
Self-testing has certainly some benefits but it has some drawbacks too :
1) It only plays against an opponent who has about the same style.
2) It only plays against an opponent who has about the same strength.

For the KID opening, this 2 drawbacks shows : It's a human who knows the strategy in this opening and who is very weaker than Stockfish.
After a lot of games in KID against a bunch of IM and GM I bet that SF would make a perf of 2800, far away from the 3200 that the small world of self-testing suggests.

Vinvin wrote:Probably because of self-testing to find improvements.
Self-testing has certainly some benefits but it has some drawbacks too :
1) It only plays against an opponent who has about the same style.
2) It only plays against an opponent who has about the same strength.

For the KID opening, this 2 drawbacks shows : It's a human who knows the strategy in this opening and who is very weaker than Stockfish.
After a lot of games in KID against a bunch of IM and GM I bet that SF would make a perf of 2800, far away from the 3200 that the small world of self-testing suggests.

Inbreeding on planet mars.

Looks like human ELO and engine ELO are different. Like comparing appels with oranges.

Henk wrote:

Vinvin wrote:Probably because of self-testing to find improvements.
Self-testing has certainly some benefits but it has some drawbacks too :
1) It only plays against an opponent who has about the same style.
2) It only plays against an opponent who has about the same strength.

For the KID opening, this 2 drawbacks shows : It's a human who knows the strategy in this opening and who is very weaker than Stockfish.
After a lot of games in KID against a bunch of IM and GM I bet that SF would make a perf of 2800, far away from the 3200 that the small world of self-testing suggests.

Inbreeding on planet mars.

Looks like human ELO and engine ELO are different. Like comparing appels with oranges.

They play the same game and they can play together so the perf rating of computers can be computed from games against humans.

Humans also have 'angstgegners'. Being a much better player but losing all the time. Like Fisher vs Geller.

https://github.com/nescitus/Rodent_II/b ... _pawns.cpp line 223 onwards. I don't claim this stuff is a readymade solution - far from it - but at least it makes Rodent's play in King's Indian a little less pathetic. At least the engine pushes the right pawns forward, even though timing is still terrible.

Theory behind that code has been elaborated by Lyudmil Tsvetkov and Carl Lagnan in the thread "Pointy chain redux part 1".

http://www.talkchess.com/forum/viewtopi ... inty+chain

I think modern engines are dangerous opponents even in the openings they play badly. But they do still have trouble with the long-term attack plans of Black in the KID.

Back a few years ago my program played some online games with some IMs and FMs and I noticed too some cases where they just gradually shrank its mobility until it was practically in zugzwang. May be harder to do now that search depths are greater and engines are stronger, but this is also a kind of long-term plan that can be hard to counter.

--Jon

jdart wrote:I think modern engines are dangerous opponents even in the openings they play badly. But they do still have trouble with the long-term attack plans of Black in the KID.

Back a few years ago my program played some online games with some IMs and FMs and I noticed too some cases where they just gradually shrank its mobility until it was practically in zugzwang. May be harder to do now that search depths are greater and engines are stronger, but this is also a kind of long-term plan that can be hard to counter.

--Jon

I think is reasonable that when engines are able to play well such schemas, they will be able also to help to discover new long term plans that people now are unaware of. And consequently, those engines will win the current ones.

There is a lot of human positional knowledge that has not been encoded in the modern chess engines yet. But the hard part is to determine which pieces of that knowledge are useful for the engines, and which are just heuristics that are needed only for humans (sub-optimal plans and patterns that are nevertheless better than playing with no plan at all). For one set of rules that kind of works for Rodent in King's Indian positions I have tested a couple of other patterns (minority attack from Carlsbad structure, a big clamp with pawns on e5 and h5 that occurs in Bf5 Caro-Cann and in a couple of French positions, a "queenside lock" with Nc5 and a pawn on a5/a4 from some King's Indians with early d4-d5),neither of which worked. On the other hand, some of the stuff that works for me looks hopelessly primitive (say, moderate amount of king tropism).

Sometimes I think that data structures commonly used in chess programs are just a bit too simplistic to allow for applying (or perhaps learning) a new knowledge. Chess programs achieved tremendous success with them, but they come from the days when both memory and processor speed were rare goods, and this heritage is clearly visible until today. I have a gut feeling that serious data mining coupled with increasing the number of piece/square tables (maybe using different sets for different central pawns structures and different king positions) might result in something as spectacular as AlphaGo neural networks. Knowledge needs a structure in which it can reside, and current chess engines don't seem to have that structure. Unfortunately creating it would require far better programmer than I am.

I know from descriptions of Andscacs that it has evolved in a program containing lots of non-standard knowledge, and I really root for Your approach to work.

Not that we know what the dev version of Houdini is doing to better handle these structures, but it's interesting that a common refrain during the Stage 3 and Rapids of TCEC was how well Houdini played its KID games.