New engine releases & news 2021

[kranium]

want elo boost? copy-paste NNUE probing code, grab strongest SF net and pat yourself on the back.
2k can become nealy 3k overnight, as has been demostrated by some.

Cheng has an excellent and unique search...adding NNUE technology need not be a sin.

You can retain Cheng's unique eval and individuality by (for ex) creating a nn eval trained NOT using SF games...
use only Cheng games...!

You'd very likely still get that 100-200 boost

[tmokonen]

what's that supposed to mean, Tony?

Yeah, re-reading my post, it might come across as some sort of accusation, but it's not. It's an expression of frustration of being leapfrogged by others in this competitive hobby of ours.

I understand your frustration, but...

Cheng didn't become 3k overnight, no NNUE either. still waiting for my magical 100 elo bugfix, it never came somehow.
my first chess program dates back to 2006. There are some who've been working for decades (Bob?) and still haven't got their well deserved 3k on single core.

kids write 100 elo stronger programs than that these days, so I must be very stupid that it takes so much time to improve.
people want real improvements (100, 200, 300, 400, 500, 600 elo) quickly, somehow I'm unable to deliver that
what to do? shrug and move on

My program dates back to 2003! I don't work steadily on it, though. I had a pretty good streak going for a while after the last release, where everything I added gave me 30+ elo, and the current dev version is approximately 250 elo stronger than the last release. The well did run dry though, and the bar is constantly raised. There's a lot of really smart people in chess programming, and it's very hard to keep up with them.

want elo boost? copy-paste NNUE probing code, grab strongest SF net and pat yourself on the back.
2k can become nealy 3k overnight, as has been demostrated by some.

I did try that. All it took was adding Daniel Shawul's library to my project, a few lines of initialization code, a for loop to fill two arrays, and a function call. It was ridiculous, a hundred easy lines of code that took only a few minutes of effort, an instant 500 elo. It was fun to see my program beat programs like Ruffian and Anmon, programs I had worshipped when I first started this hobby, but it wasn't very satisfying because I just bolted on someone else's code and net. It was essentially making my program Stockfish with a 2150 level search and a slow movegen. My next release will be open source, once I clean up some supremely ugly code (you do NOT want to see the inelegant way I handle en passant), but I won't distribute an NNUE executable. That I will leave for someone else to compile.

PS still waiting for a CCRL rating! (and I'm not someone who releases each week or two)

They did test 4.41, but only a few games on a single CPU. I am not sure why they didn't test on 4 CPUs, as 4.41 is clearly stronger than 4.39. It deserves to be tested on 4 CPUs. If I see Graham in TLCV chat I will gently remind him of this glaring omission.

Anyway, this thread is supposed to be new engine announcements, so I won't clutter it further with this topic.

[amanjpro]

PS still waiting for a CCRL rating! (and I'm not someone who releases each week or two)

I'm someone who does cut a release every other week, more or less. And I don't think this is bad, I mean I have an elo boost why shouldn't I share it with everyone as soon as possible, that means more versions for our dear testers to test. But this is not necessarily a bad thing, as long as the version brings some value.

When I was trying to find new opponents for Zahak to test it, I had a very very hard time to find an open source engine, that can easily compile on Mac and is (obvious/big) bug free and is in the same range as Zahak (a fair test case).

To make it easier for future devs, I try to release Zahak often enough to fill this gap, every 100 elo boost or two is a release so the next dev has something to use for his tests

[mar]

I'm someone who does cut a release every other week, more or less. And I don't think this is bad, I mean I have an elo boost why shouldn't I share it with everyone as soon as possible, that means more versions for our dear testers to test. But this is not necessarily a bad thing, as long as the version brings some value.

When I was trying to find new opponents for Zahak to test it, I had a very very hard time to find an open source engine, that can easily compile on Mac and is (obvious/big) bug free and is in the same range as Zahak (a fair test case).

To make it easier for future devs, I try to release Zahak often enough to fill this gap, every 100 elo boost or two is a release so the next dev has something to use for his tests

sure, a big jump is an exception, I was thinking about things like +10 in self-play or bugfixes that don't really fix anything.
also you need to understand that you cannot keep gaining 100 elo in a week or two indefinitely (hint: the higher you get, the harder it is to improve)

I'm not sure if testers are very happy with frequent releases: imagine you've spent a lot of time testing version A and before you even get a rating (say 1k games) version B is already out (I hope the point is clear)

[Modern Times]

PS still waiting for a CCRL rating! (and I'm not someone who releases each week or two)

The FRC rating has been there since around 5th May.

http://ccrl.chessdom.com/ccrl/404FRC/cg ... _length=30

[Gabor Szots]

PS still waiting for a CCRL rating! (and I'm not someone who releases each week or two)

Hi Martin,

Let me explain my policy with Cheng. From what was claimed I gathered that 4.41's progress compared to 4.40 will be relatively small, so small that it would be suppressed by the error margins. Therefore I decided against testing it the usual way (by running a long gauntlet), instead I decided to replace 4.40 with 4.41 in all my tournaments so that 4.41 will slowly build a rating. This is why 4.41 has only 66 games in our blitz database.

I have done this already a couple of times with other engines as well. This frees up some CPU resources for being able to test more engines. Later, if I see that my estimation was wrong I change my plan and start a gauntlet to accelerate the process of obtaining a new rating.

Those 66 games show that 4.41 might require its own gauntlet for testing but I'm not very sure although Ray's FRC results confirm that the improvement might be bigger than I expected. If your expectation is in line with those results please indicate so and I'm going to promoto Cheng in my testing queue.

[Gabor Szots]

I'm not sure if testers are very happy with frequent releases: imagine you've spent a lot of time testing version A and before you even get a rating (say 1k games) version B is already out (I hope the point is clear)

True. In such cases my solution is to skip version B.

[mar]

@Ray: a poor choice of words for which I apologize. I know I got a FRC rating quickly for which I'm very grateful
Gabor, let me explain (and then I'm bailing out because I'm already starting to pollute this thread which goes against the promise I made to Guenther

I measured 56-57% in self-play against 4.40 at very fast TC (10+0.1 and 60+0.6 sec). I usually get less than half of that (in CCRL 40/15 for sure)
The problem is how to extrapolate? I simply gave a conservative (hopefully lower bound) estimate, because I didn't want to lie about the improvement (unlike some people who make ridiculous claims and then play surprised - so I believe it's better to give a lower estimate)

With 4.40 the rating spread among various rating lists was over 40 elo (!) which is a lot for my taste. the outlier being CCRL 40/15 where I got like +35, +60 in CEGT 40/20 and CCRL 40/2 IIRC, over 80 in FastGM and CCRL FRC. at this level, that a huge discrepancy.
For example, Tucano 9 seems underrated in CCRL 40/15, Amoeba 3.3 (which show a clear improvement in CEGT 40/20) is "worse" in CCRL 40/15 after 700 games.

could be error bars, could be the openings, could be the TC, the choice of opponents or even some problem with the engine, honestly I don't know

CEGT 40/20 certainly shows more than my +15 guesstimate (around 50, which will likely drop as more games are played), so I still have hope that it's more than I expected.

while self-play is a big gamble (you can get anything between 25-90% of what you measure), it's still the way to test small elo changes with many games. considering the huge spread among rating lists, I think it's virtually impossible to give any reliable predictions until the engine is actually tested by independent testers. (and unless the improvement is like 150+ elo in self-play, which is a lot at this level)

[AlexChess]

REAL ENGINES ELO COMPARED TO HUMANS?

I have asked this question to the mythical Crafty's programmer Dr. Bob Hyatt 20 years ago https://www.chessprogramming.org/Robert_Hyatt , when best engines were 2400 ELO and GMs were easily winners against them. Now GM Hikaru Nakamura has asked 2 more pawns to accept a blitz match against Dragon 1.0 by Komodo https://www.chess.com/news/view/hikaru- ... ess-engine . Personally I think that 3600 ELO rating calculated by some famous rating lists (for Stockfish 13-dev) is an optiminstic evaluation even with a Ryzen 5900x with 2 or more Nvidia GeForce RTX 3060 graphic boards. I think that 3200 ELO BLITZ in more realistic for top engines, 3000 ELO on long time games.

What is your opinion?

Regards, AlexChess

[Damir]

CrazyAra & ClassicAra 0.9.3
https://github.com/QueensGambit/CrazyAr ... /tag/0.9.3