lichess.org

Large Atomic openings study

As I mentioned in the Spam Game topic (lichess.org/forum/team-atomic-chess-theoreticians/beat-the-spam-forum-game ), I started to build a hand-made opening book (that'll probably becomes in the future the base for an engine-evalutated book): lichess.org/study/HQ3D3Cme It will have a chapter for every major Atomic opening, and at the moment I made a draft for 1. Nf3 f6 2. e3 as the most popular opening. What it's definitely lacking for now is 3... Bg4 lines, which I add very soon, but I also want to get some feedback if what I've done so far is to any degree adequate. In particular, I'm interested if
1) There are some serious mistakes such as the line that's marked as drawn is known to be not drawn at all, because [the line]. (Or the reverse)
2) There are some <s>annoying</s> interesting lines whose refutation is not shown.
3) There are some non-shown refutations which are known to be much more solid than shown.
By the triangles I marked the positions which I consider to be "theoretically important". For all of them the minimax evaluation of leaf nodes is set. I also appreciate the feedback if I need to add more triangles or remove some etc.

Of course, the evaluations are largely subjective. Ideally, in controversial cases I'd like to discuss the line first, and if the issue still remains unresolved after the discussion - play an engine-assisted game if the line seems pretty relevant. But at the moment I haven't elaborate on the precise rules and just want to know your opinion on if something like this is helpful at all.

In the background auto book builder is running day after day. Currently it has cca. 9000 posiitons analyzed to depth 15. In a sense it can be viewed as a large opening study.

Some relatively interesting tendencies can are already be observed:

https://imgur.com/Hx7xb7i

In the magenta column you can see the number of nodes behind each move as counted during the minimax process.

1.Nh3, 1.d4 and 1.e4 have astronomical node numbers, while 1. Nf3 and 1. e3 have low node numbers, you would normally expect based on the total number of analyzed positions.

This is due to 1.Nh3, 1.d4 and 1.e4 leading to slow, maneuvering positions, where move order does not matter that much, so you can reach same positions through different move orders, and all the nodes in those lines are counted. This also points to the difficulty of auto book building in general: these projects seem to die a minimax death after a certain point. We are not there yet, but minimaxing the tree takes on the order of 10 minutes already.

Also 1.Nh3, 1.d4 and 1.e4 seem essentially the same moves, because their evaluation is the same, suggesting that no matter which of these moves you make first, sooner or later you will make the other too and get essentially the same positions. These moves are interlinked and form a system.

With 1.Nf3 and 1.e3 the game is forcing, there is only one way forward, hence the low number of nodes. I personally think, these two moves are stronger than the other three. I also think that apart from these five moves, white has no other potential winning moves.

Well, I'm not sure about 1. d4, but it's quite weird that 1. e4 gets to the same positions that 1. Nh3. After 1... e6 2. d4 local Stockfish suggests 2... Nh6 which looks logical - why allow white's knight develop first? abb suggests 2... d5

1.Nf3 f6 2.Nc3 Nh6 lines seem to go to +4 evaluations at depth 40. May be I'm missing a move somewhere, but fighting such evals without much hope for pawnitized endgames is not lucrative.

Instead I added 3..d5 to 1.Nf3 f6 2.Nc3 c6 3. d4 ( in practical play 3..e6 fares better, this is a candidate if 3..d5 fails). The main line 3.e3 e6 has already been played, so as black I have nothing to do there.

1. Nf3 f6 2. Nc3 c6 3. d4 d5 4. e4 Bg4 5. Nb5 dxe4 6. f3 cxb5 7. Bb5+ Nc6 8. d5 e6 9. d6 Qa5+ 10. b4 Kd8 11. bxa5 Bf5 12. Be3 seems to be lost, and with it 3..d5 is very likely refuted.

This is a real shock, it means that neither Nh6, nor c6 can hold the two knights attack and it converges to victory more quickly than atomic attack with e6.

As engines get stronger and new people come into the game, theory has to be reevaluated.

I think these new ways of playing the two knights attack by Wolfram_EP are great additions to atomic theory.

It would require an original player like Xeransis to try to come up with new ways of defending for black against this kind of play, if you just look at alternative engine moves, they are worse than defenses tried so far.

Although this is indeed interesting from the theoretical viewpoint, I'm not sure this makes the whole game much easier to win in Blitz games. Sure it works quite good against strong players who aren't specificially prepared to play it, but in fact there are many sidelines, it is enough to prepare one of them quite well as black to a match, and practical chances are more or less equalized, because white must play practically the same moves till large depth being prepared to any black deviation, and black can randomly choose a sideline at any point. The resulting positions of these sidelines, despite being objectively something like +4-+5, are quite hard to get right for a weakly prepared human - at least for me:) (and white will be less prepared than black in a concrete line).

I played something like 20 games against @AtomicChessBot using this system trying to squeeze it and just lost all of them, because it is such a slippery. Obviously my current level of preparation, despite I played quite a lot of games vs humans there, is not even remotely satisfactory for a serious match. You need to, like, memorize lines systematically, and not by routinely playing humans, because many decent lines are not played (almost, or at all).

@linetester3

For to say that this line is a win, I think maybe you are not looking at enough options. 3... e6, 4... Nh6 or maybe even 8... Nh6, 8... e5, 11... Ne5 all are look interesting. Not even does look such bad the final position with 12... Nb4 13. c4 Nc2+ 14. Kf1 Rc8 or 14... h5 from a viewpoint of practical, even if it is a loss of theory. Most interesting I think is certainly 4... Nh6 5. g4 e5! 6. Nb5 Bb4+ 7. c3 O-O 8. Nh4 exd4 9. f3 where not I am sure it is such easy to say what is should be best move for black of theoretical of 9... g6, 9... Bc5 and 9... Na6. Maybe it is bad for black of these positions, but of certain that not stockfish can say that it is such simple and is even more hard of for humans. I think that it is better for white to play 3. e3 instead of 3. d4 to avoid such complications and to have position that is almost certain a win.

I think maybe it is needed for players to stop believing that stockfish is always correct and instead look at how complex is actual position.

In antichess some winning lines are so absurdly counter intuitive and require memorization of hundreds if not thousands of lines that white in some cases is better advised to play some objectively drawn opening moves which nevertheless provide huge practical advantage with black virtually having to make only moves to save the day.

The same is not true of atomic chess. Atomic chess is much more intuitive and for white playing best known theory is almost always the answer.

I don't believe that Stockfish is flawless, but I do believe that 99% of positions which Stockfish values +4 at depth 40 are wins. The only known systematic flaw of Stockfish is not understanding pawnitized bishop endgames, but this can be filtered out using human assistance. All positions Wolfram_EP claimed to be wins and some of which we finished analyzing at +4 depth 40 cannot possibly be reduced to pawnitized bishop endgames.

All in all I think this work done by Wolfram_EP is important not only theoretically but practically, and he is certainly a much better theoretician than practical player.

If you feel you can think of moves that can at least can complicate things for Stockfish, than go ahead and make them in our theoretical game:

cserv.herokuapp.com

@linetester3 Stockfish also doesn't understand endgames rook vs bishop with even pawns without passers, some of which are drawn.

Reconnecting