Can black draw?

@Nordlandia

I think the more accurate reading is that the top 3 engines lose this as black :)

In all seriousness, I wouldn't be surprised if it turned out white has a study-like win, nor would I be surprised if it turned out to be a draw.

Prior to any serious analysis, my instinct is that this should be a draw. The problem engines have is that this initial position is given an evaluation in a range where almost every position with such an evaluation is lost.

When you get an edge case that happens to be a draw, it's easy for an engine to go from a drawn -1.7 to a lost -1.65 or such because it thinks the lost position looks better in the relatively short run of its calculations.

It's an interesting position...I'll probably try my hand at some analysis of it and see what conclusions I can draw.

Short answer :
If Sarg0n and Nordlandia (mere 400 points ahead of me :) ) don't know, I'm in no position to say otherwise. Play on.

Long answer :
Capablanca 's standard plan when you've won the exchange is to give back the exchange to win a pawn in a good position (i.e. in a position where the extra pawn is indeed winning).
In this position, if you give back the exchange on Black's knight, you end up with opposite color bishops. One extra pawn is far from enough to win in principle. If you give back the exchange on Black's bishop, you end up with a bad bishop vs a good knight. Again, one extra pawn is not enough to win in this case. So the standard plan is not available here.

You will need to play this as a middlegame, not an endgame, and make your extra rook count (more than Black's extra minor piece). For this, you need to open up a column while keeping pawns on both sides. There is only one candidate for that : the f-file (or g-file, depending on Black's reaction). White must play g4 and f5. But, Black can leave pawns on only one side by pushing h5 first (with White to play, 1.g4?? fails horribly). After the exchange of h pawns (and maybe g pawns), Black has countered White's plan, but the f-pawn is weak.
White should now try to win Black's f-pawn, probably by posting the bishop on the a3-f8 diagonal and by threatening repeatedly to exchange one pair of rooks (Black doesn't want to exchange rooks, as stated by Nordlandia). I think that Black will consider playing f6 on his own to prevent that. Also, the h5 push had a downside : White can push h4 and an exchange hxg4 would then create an outside passed pawn (h).

As long as you have an A-plan and a B-plan for White, Black's defense is not trivial. The position is both hard to break and hard to defend. Mistakes can occur on both sides. So my conclusion is : I do not know.

I've put this position on Komodo (where I saw Nordlandia's analysis as well :) ) and here is how White makes some progress :
15.Ke2 h5 16.h4 Rf8 17.f3 f6 18.exf6 gxf6 19.Bd2 Rg8 20.Rag1 Ne7 21.Rh3 Bd5 22.b3 Nf5 23.Kf2 and either 23...Nxd4 24.Bc3 Nf5 25.Bxf6 or 23...Rc8 24.Rc1 Rg8 25.Ba5 opens a second frontline. This is still difficult to win, but the "two-weakness-principle" might apply. Black's drawing chances seem to be good in the second line : 25...Kb8 26.Rhh1 Bc6 27.Rcg1 Bd5 is Komodo's suggestion. The assessment is still +1.6, winning for White, but White makes no further progress (Komodo shuffles the rooks along the first rank).

A very nice and instructive position to analyze !

Here's a good example of what I mean.

I like to start such analyses by playing some games against the engines, trying various shuffling setups.

Here's my second game against asmFish at 3 seconds per move (the first one I did my usual first step and tried to literally just shuffle without changing the pawn structure just to see what plan SF would come up with; h4-h5-h6, f2-f3,g2-g4,f3-f4-f5 was its brute approach which turned out to be rather effective):



Now, if you have the patience to sit through SF's "brilliant" attempts to win, that illustrates my earlier point pretty well.

The engine screamed bloody murder at some of my moves, and thought it was getting a much bigger edge than the "optimal" lines, which typically involved opening files for the rooks.

Given what I saw of its variations, I'm not surprised that the engines end up losing this as black. A lot of variations are like the one @A-Cielbleu posted, where black ends up opening some lines.

Even some of those positions may be drawn, but there's no reason to help white out like that.

Of course, on bigger hardware and/or with more time, it's possible SF's play in my game is improved (some of its moves, especially a5?, were just silly, but then again, giving my low level of chess skill, I'm sure so could black's).

Either way, it illustrates my point pretty well. The engines have a hard time because they see that in these sorts of lines White ends up with a lot of space, and black is basically down a whole rook since it's trapped on h7, so they give big eval bonuses for white, and avoid the lines as black, even if they may be the only sorts of lines that draw.

Conversely, when playing the attacking side, the engines are easily baited by these sorts of positions, and don't always try the most testing approaches.

I'll be doing some actual analysis soon, but I was surprised that SF let me draw so easily, so I thought it was worth a quick update :)

Interesting. It shows once more how weak Stockfish is in endgames without its table bases.
White should place the bishop at d2 to restrict the black knight, put pawns on white squares, open not close the position and work to create a passed pawn. The rook is only stronger than the knight if the rooks becomes active. So white must aim for open files and for a passed pawn. In a closed position the defended central knight is no weaker than a rook.

Commentator Maurice Ashley said yesterday that if the side up the exchange don't have open lines for the rook, the extra exchange is virtually irrelevant.

@tpr

Well, to be fair, while it definitely played both sides of this position relatively poorly in my first few quick games, this is very much an edge case.

For every position like this, there are thousands of positions it plays very, very well.

In fact, the general consensus among strong correspondence players and advanced chess players who hang out at places like TCEC chat, talkchess.com, and rybkaforum.net is that SF plays endings very very well.

I'd agree with that in general. There are positions like this where it makes itself look a bit silly, but they're not nearly common enough to be an indicator of general level of play.

Interestingly, I also ran through that game with Critter, which while an older engine at this point is somewhat unique of the very strong engines by having some built-in attempts at fortress detection.

It works pretty decently, and lets Critter notice some draws that SF, K, and H still don't quite evaluate as drawn.

It's not enough in this game, though; even after the position was completely locked it still gives large evals for white.

I'm still trying different attacking and defending setups, but so far it's looking very much like a draw. I've found several defensive setups SF hasn't been able to crack, but no attacking setups that panned out.

If there's a win, it's probably cleverly hidden :)

I'm convinced that a completely passive defense by Black is not going to hold.
15.Ke2 h5 16.h4 g6 17.f3 Rh7 18.Bd2 Rh8 19.b3 Rh7 was played in Stockfish-"a_pleasant_illusion" as a fortress example, and there I suggest 20.Rag1 Rh8 21.g4 Ne7 22.Rh2 Rh7 23.gxh5 Rxh5 24.Rg5 and Black must allow either a trade of rooks or an advance of the h-pawn. In the first case, White plans Be3-f2 and Rg2-g8-f8(h8). In the second case, Black's rook will be more passive than any of the white rooks.
There is no way Black can keep the kingside closed for ever.

I think f6 (answering exf6 with gxf6) and active defense (Bd5/Rc8-g8) is the path to a draw, if there is one. Yesterday night I tried a g-pawn push, letting the black rook invade on c2. Trading the black knight and winning the black bishop against a lot of pawns (by promoting the g-pawn), I ended up with two rooks against rook and four well-placed pawns and I couldn't make progress there.

This analysis is definitely beneficial for my chess. Thanks to the original contributor Nordlandia !

@A-Cielbleu

Yes, that is one natural attempt SF and I debated for a while.

Even once that's accomplished wins are not particularly easy to demonstrate, but there's a very good chance that white wins all variations there.

That same plan doesn't work very well against another passive setup, though (same thing, but holding off on g6 until needed, instead playing an early a5 so the queenside can be locked).

The active defense plans do also look promising from the dozen or so games I've played with them.

Anyway, as I mentioned in that post, the claim/point wasn't that I thought that particular shuffling setup forced the draw.

The game was illustrating my previous point that it's hard to trust the results of self-play of the top engines in such near-fortress positions, because engines still sometimes play both the attacking and defending sides poorly because of misevaluations of the actual fortress positions that arise in some variations and must be sought by the defender and avoided by the attacker.

The fortress I obtained and SF allowed is just one such example.

@a_pleasant_illusion TCEC is with table bases and then Stockfish plays endings much better. Even in TCEC Stockfish seems better at drawing than at winning now.
I agree with @A-Cielbleu. White must place his pawns on light squares, not on black squares and white must never allow pawn chains to close.

@A-Cielbleu yes, black can not stand idle.