K+2N vs. K+P

https://syzygy-tables.info/?fen=4k3/3n4/8/8/8/8/3BB3/4K3_w_-_-_0_1

26.8% KBB wins

13.5% KBB wins without the 50 move rule

59.7% Draw

0.0% (532 games) KN wins

Yes those data are extremely mind blowing. 

Look at two diagrams.

Diagram 1 is draw.

Diagram 2 , white win.

No one on earth can tell those outcomes.

If you run stockfish at your home computer, it will takes at least several weeks to find out those results.  

Well I could tell you those outcomes at a glance, as could many other people.

After 1...b5 in the first diagram there's no win, 1.Nc4 c5 is a standard win in the second. Troitzky analyzed that before computers ever played chess. 

Far too much emphasis on computers these days.

I think you may be very optimistic to expect to find out those results by running SF on your home computer.

I tried setting up both positions in SF and started the Kibbitzer in each and left them running for three hours checking each after each hour. I assume that's essentially the process you envisage, but over a longer timespan.

The draw showed a score of 6.40 after each of the hours and the win a score of 6.75.

How are you ever to know you have a draw? I would expect the draw to continue showing something only marginally different ad nauseam. I don't think the 50 move rule is taken into account by the Kibbitzer ( though repetition probably is).

On the other hand looking at the timing and play I would also expect the same to be true for the won ending. SF gives four lines but its preference order as White doesn't match the closeness to mate achieved. It starts with mate in 38 and the best it achieves after 15 moves (the full three hours) is its third preference at mate in 33. On average it achieves one step backwards, i.e. mate in 39. Looking through the play it suggests a random walk in terms of mate distance.

That means that although although on its best attempt it has mate in another 65 ply the expected depth of the search would actually be the expected number of steps of the random walk to reach -65 on the distance to mate.

Worse the time taken to add a ply to the search depth appears to multiply by about 10 in consecutive hours and is down to about 1 per hour at the end on the three hours. (I suspect that ratio could actually get worse as the search progresses).

So as a conservative (but flagrantly unscientific) lower bound for the time needed to complete the search search I'd guess 10 to the power 65² hours or about 10⁴²²⁰  years. Your home computer may be warranted for that long, but I don't think the Universe is. 

An interesting question. The fact that it can't apparently distinguish between won and drawn positions would suggest not.

I'm not set up to try it. It needs an interface that will allow native Stockfish as one player and a Nalimov attached engine as the other. Do you have such a thing?

Otherwise I could try running both and transferring moves between boards, but that's rather a lot of transfers. Or I could just try playing Black myself and see how it does. 

If you can run it you probably need to give SF a reasonable amount of thinking time, say G/120+10, it's very much affected by that. 

Edit: I just tried playing SF both sides from the won position and SF seems to play it pretty well either side. I played it with the above time control. Presumably it Kibbitzes badly because it's using insufficient time for its moves.

Second edit: @Prometheus_Fuschs: I tried playing Black in the won position again and I've changed my mind. The first impression was just showing my rustiness in this flavour of KNNKP. At the second attempt SF lost as White on time with standard time controls, but had no chance of winning within the 50 move rule anyway. I also found it can't play the White side of rooks pawn endings at all.

Endgame table bases are perfect , no one can win against tablebase- 0 chance

Stockfish is not perfect, stockfish will always lose against tablebase.

I think SF will cede moves as either side against the Nalimov tablebase in your won position, but I believe it would still win the position comfortably within the 50 move rule (given an adequate time control).

Edit: I take that back - see post #67.

When you say nobody can win against the tablebase, obviously people can win from won positions. Syzygy will beat Nalimov in some positions with the 50 move rule in effect when Nalimov may not win against itself and Nalimov will effect quicker mates or more robust defences in some positions with or without the 50  move rule in effect. It is true for the position you show that nobody will mate the EGTBs in less than the 38 moves it should take, against either of these, or defend longer.

Your definition of perfect moves is probably different from mine given the fact that I can win against any tablebase in simple endgames (KvKR for example).

Can you give an example?

@Prometheus_Fuschs: That should be enough - played at G/120+10. 

I must admit I understood incorrectly, I thought about Rook and Pawn endgames which puzzled me since they are rarely a subject to study with TBs.

Thank god we can implement tablebases in engines though.

OK if you can get that far against SF.

I practice general endgames against Nalimov because it's quicker and more accurate. I used to have a version of Rybka with an "e" (for endgame I think) on the end which was reasonable on KNNKP but it's currently on a defunct computer and the new version is totally useless. 

I do occasionally practice that one against Nalimov too. I can pretty well guarantee to reach mate in the right number of moves from any position playing either side. 

Hmmm, does Nalimov allow you to play against itself like a macth game? I have 6 men Syzygy on my computer but I only use them for the engines.

I use chess program Wilhelm with Nalimov attached. You can set it to auto respond when you make a move or click the move button. It will also give you a single line Nalimov v Nalimov.

Wilhelm is useful because you can ask it for positions with a range of moves to mate and fix some or all of the pieces and set relative positions. It will also give you things like the win/draw zones in KNNP for the Black king for postions of the pawn beyond the Troitsky line or the zones when the pawn is blocked by a knight.

Nalimov doesn't take the 50 move rule into account, but I don't think I'll ever try to learn how to play taking the 50 move rule into account in any endgame where it's relevant. I've only got the 3-4-5 man Nalimovs at the moment.