A 3000 could easily beat a 2000, but could a 4000 easily beat a 3000?

From Github SF overview.

If the engine is given a position to search that is in the tablebases, it will use the tablebases at the beginning of the search to preselect all good moves, i.e. all moves that preserve the win or preserve the draw while taking into account the 50-move rule. It will then perform a search only on those moves. The engine will not move immediately, unless there is only a single good move. 

But as I said before, I'm interested in native performance without an EGTB.

What I mean the Search is Alpha Beta Pruning Search of Stockfish.

The one it is written in yours is Tablebase Probing.

And yes, Stockfish stop AB pruning serarch and choose the moves according to Tablebase probing. 

See, GUI setting in my Stockfish. 

@drmrboss

My interest being in native strength (no EGTBs), this would be governed by the normal search. Presumably the search mentioned in the quote I gave is doing something like finding paths with the smallest fraction of perfect responses and may have remained unchanged through SF generations (I don't know).

At any rate SF12 is no good for my purpose because it can't play KNNKP native. I got a version of Rybka with an "e" for endgame on the end of its version number that could do 50 move mates in KNNKP under its own steam about 15 years ago, but carelessly lost it on a defunct disc.

Do you know if LC0 with GPU can play the endgame under its own steam? I'm interested in playing it myself rather than watching two engines play each other and I can already practise against Nalimov.

Mates much beyond 50 moves in KNNKP tended to be messed up by Rybka because of the 50 move rule handling which as far as I know was not switch offable. It should be these days because it's no longer in the basic rules, but I don't know if it can be in Lc0. Do you know?

hello

I do not know if people can reach 4000

Some tablebase positions requires a few seconds to minutes per move to solve 100% correctly. But some tablebase positions will requires months to years  CPU power to solve 100% accurately. That is why a chinese guy donated 7 men syzgy tablebases to public to save massive computing powers.

I highly doubt a 15 years old engine is better than modern engine unless you are trolling and wasting my time. But hey, it is a free world, you can believe whatever you believe. You can believe a 15 years old Nokia 2G phone is better than current  5G phones. 

It is more possible to construct a program that takes into account the longest, most challenging route to a draw in a known drawn position than it is to create a 32 piece table base in the first place. 

A perfect engine will take into account the proclivities of its opponents simply by reviewing their extant games.

No trolling; it's correct.

Rybka "e" on the end version could mate in KNNKP depth 50 positions without access to a tablebase. I think the last "e" on the end version was about 15 years ago.

SF12 can't.

Neither can the latest version of Rybka, or at least the version I downloaded which is probably not now latest. (It can't even manage KBNK, in common with the GPU free LC0. GNU chess could manage KBNK somewhere around 2005.)

I suspected at the time that the program contained specific code for KNNKP, though I don't know. It wasn't excessively accurate, but it could force mate.

(And as far as 'phones go, I much preferred the ones that stayed at home on the sideboard with the dial on the front.)

So there are engines that can calculate those elusive endgames without using a tablebase?

Rybka with the "e" could, but it's the only one I've come across. It started to fall apart when the depth exceeded about 55. I think that was the 50 move rule code putting it off its stride.

No.

Depends on complexity of position, engines will requires from millions of position search to Trillions or Decillions of position search to play 100% accurately. For example, 7 men TB with 524 moves checkmate will takes several years of computing power in your home computer to solve it.

Current engines like Stockfish 12 are ways faster than 15 years old engine. If you test 100 endgame positions, I will say SF 12 can still get ways higher success rate than a 15 years old engine.

@endgamestudier, if you really would like to get technical details of chess engines, I suggest you join talkchess forum, whereas 80-90% of top engine programmers discuss those stuffs.

Dude, stop telling non sense.

Bring 100 random positions of KNN KP or whatever 100 FENs you like. 

Ask your favourite engine play 100 games himself in whatever time control your like. Check the accuracy with Tablebase.

Do the same with Stockfish 12 or Stockfish 13 developmental version ( Disable NNUE).

Post PGNs with depth and nodes per move data attached and statistical results here as a proof.

Why I ask evidence? Extra ordinary claim need to proof with extra ordiary evidence. Another sensible logic, if you claim your granddad's 50 years old car is faster/ better than modern cars, nobody will believe you unless you show reasonable proof/videos.

Just look at post #379. That is SF12.

It's supposed to be mate in 50 but all SF12 can manage is a draw in 8.

I played Rybka with "e" on the end in many mate in 50 positions on slower kit and it could always manage the mates - though it might have needed the 50 move count reset. (Can't prove this any more because I've lost the program. That's why I'm looking for a new one.)

I'm sure SF12  runs much faster, but when it attempts KNNKP under it's own steam it's obviously running in the wrong direction.

You need to do at least 100 positions ( or minimal of 20 before starting conclusion).

A single game , a single result mean nothing.

e.g GM Kramnik missing mate in 1 should not be assumed as 300 or beginner. Look at 100 games of GM Kramnik.

Why do I need 100 games? I played 47 moves against SF12 in the same example and it ceded me 5 moves. With that accuracy rate it will not generally manage a 50 move mate whatever the position. 

If Kramnik were to miss 5 mates in 1 out of 47, you could indeed conclude he's not very good.

I have actually played at least a dozen similar positions against SF8 and it couldn't play any of them.

It is not a rule, but larger sample always give you reliability.

Otherwise, you get that hilarious result that no statistician/scientist believe.

Then you may ask why people from Stockfish test 40,000 to 60,000 games of regression tests,

Simple, they dont want to conclude hilarious result like you. 

@drmrboss

Well I'm not going to play 100 mate in 50 positions at G120+3 for you. They take about 40 mins each. You'll just have to stay deluded.

 4:3 is smaller than 3:2, so the 4000 would still have an edge, but definitely not as one-sided a result as that of 3000 vs 2000.

As a ratio, you'd be right. But this isn't about ratios, it's about the distance between the two points. At any two points on the number line, if the difference is the same, (4000-3000, 1500-500, and so on) then the skill difference (And expected results) are the same. It's not a ratio. It's a number line.

But that's just it. Some rating systems follow square root formulas in their calculations, so unsure how linear it really is. Let's try and keep this thread civil btw, one of the best threads, don't wnat locked!