Ceres Chess Engine

mehmet1921 · Post by **mehmet1921** » 01/01/2021, 5:32

Ceres (MCTS chess engine for research and recreation)

Ceres (Chess Engine for Research) is:

A state-of-the-art UCI-compliant chess engine employing the AlphaZero-style Monte Carlo Tree Search and deep neural networks
A flexible, modular and efficient software library with an exposed API to facilitate research in computer chess
A set of integrated tools for chess research (e.g for position analysis, suite testing, tournament manager)
The Ceres MCTS engine is a novel implementation written in C# for the Microsoft .NET 5 framework.
This system comprises about 50,000 lines of source in 400 source code files, developed as a way to try to make something good come of COVID confinement. The underlying neural networks (and backend code to execute them) and backend code are currently mostly borrowed from the LeelaChessZero project via a "plug-in" architecture.

It is important to acknowledge that this project stands "on the shoulders of giants" - the pioneers in the field such as DeepMind with their AlphaZero project and the ongoing Leela Chess Zero project.

For more information about Ceres:
https://github.com/dje-dev/Ceres

mehmet1921 · Post by **mehmet1921** » 01/01/2021, 5:45

Test made by dje (Lco Discord Channel)

Match of LC0 vs. Ceres (a new MCTS engine leveraging the LC0 backend)
LC0 version: lc0-v0.26.3, backend=cuda-fp16
LC0 options: 703810 with nncache=5000000, deumux backend (2 GPUs)
Ceres version: 0.8
Ceres options: default (also with 703810)
Hardware: dual Intel 6142 CPU with 2 A100 GPU
Time control: 120s+0.5s
Book: TCEC19
Tablebases: Syzygy 5+6 man
Adjudication: tablebases or 4 consecutive play both 350cp
Software: Ceres integrated tournament manager
Comment: These first published results for Ceres require independent community verification.
Comment: The realized NPS ratio of Ceres vs LC0 was 2.3 in this test.
Comment: Ceres strength is massively exaggerated in this unrealistic combination of small nets and big hardware.
Comment: Ceres will be released under GPL3 on Github on Dec 31, 2020 (today).

Tournament Results of Ceres1 versus LC0_0 in 102 games
Ceres1 wins: 33// Ceres1 draws 69// Ceres1 loses 0

# PLAYER : RATING ERROR POINTS PLAYED (%) CFS(%) W D L D(%)
1 Ceres1 : 0.0 ..... 67.5 102 66 100 33 69 0 68
2 LC0_0 : -128.6 .. 38.4 34.5 102 34 --- 0 69 33 68

White advantage = 107.97 +/- 20.04

kramnik · Post by **kramnik** » 01/01/2021, 8:48

Are the executables hosted somewhere? I just can see the source here and installing Visual Studio and learning how to compile this engine just for these executables is a pain. I'd be curious to test it on a longer time against some Stockfish-derived engine.

Nemesis · Post by **Nemesis** » 01/01/2021, 8:53

kramnik wrote:Are the executables hosted somewhere? I just can see the source here and installing Visual Studio and learning how to compile this engine just for these executables is a pain. I'd be curious to test it on a longer time against some Stockfish-derived engine.

Binaries not available yet - I agree - pain in the butt compiling.
I'll upload binaries unless somebody else uploads before me.

Nemesis · Post by **Nemesis** » 02/01/2021, 9:19

We're getting there .. slowly .. I have NOT tried this and probably wont.
Posted this for info on progression.

ribbit Yesterday at 5:49 PM
A very short summary of bullet-points of the quickest way to get going with ceres on a win 10 box (have only tried on Intel cpu)

• Download Lc0 26.3

• Download .Net 5.0 & install
https://dotnet.microsoft.com/download/dotnet/thank-you/runtime-desktop-5.0.1-windows-x64-installer

• download Cuda 11.2 from
https://developer.nvidia.com/cuda-downloads?target_os=Windows&target_arch=x86_64 and install.. ensure path and CUDA_PATH point to the 11.2 directory.

• get the boxed version of ceres and put it in the 26.3 lc0 directory
https://www.chadhosting.xyz/programs/Ceres_boxed.exe

• get the ceres.json file place in the same directory 26.3 directory
https://cdn.discordapp.com/attachments/427066771627966466/794298546199068692/Ceres.json

• download lc0.dll (non mkl version)
https://cdn.discordapp.com/attachments/427066771627966466/794291040983056414/lc0.dll

• create a sub-directory to where you placed the ceres_boxed.exe named /runtimes/win-x64/native place the lc0.dll in the "native" directory

• Start ceres_boxed.exe and run go nodes 10 uci command to check.

Nemesis · Post by **Nemesis** » 02/01/2021, 9:48

@MCDermont :

Your post has been deleted - duplication of mine.

MCDermont · Post by **MCDermont** » 02/01/2021, 9:50

janus wrote:

@MCDermont :

Your post has been deleted - duplication of mine.

sorry i saw yours to late and wanted only help :-)

LittleJoe · Post by **LittleJoe** » 02/01/2021, 12:16

More on talkchess: http://talkchess.com/forum3/viewtopic.php?f=2&t=76201

Nemesis · Post by **Nemesis** » 02/01/2021, 13:24

I followed the instructions posted above by 'ribbit' at Discord and I have a working configuration via the command line.
It doesnt seem to work under Nibbler so I'll need to ask at Discord.

EDIT >> works perfectly using Arena and Shredder GUI's.

Nemesis · Post by **Nemesis** » 02/01/2021, 17:51

From TalkChess ..

mehmet1921 · Post by **mehmet1921** » 03/01/2021, 3:59

Test by Moisonov ("Clementine"//Lc0 Discord)

Match: Lc0 26.3 and CERES vs Stockfish 201225 bmi2 (18T)
Time control: 120s + 2s
Hardware: RTX 3080, i9-10900kf
Openings: TCEC SuFi's 10-19 (500 pos.) with a few modifications.
Stockfish: threads=18, hash=2048
Lc0: cuda-fp16, nncache=10,000,000 ,threads=2, minibatch-size=204, mlh=on
mlh tcec-19: moves-left-max-effect=0.2, moves-left-threshold=0, moves-left-slope=0.004, moves-left-scaled-factor=1, moves-left-quadratic-factor=0, moves-left-constant-factor=0
TBs and adj.: syzygy 5-men, resign movecount=5 score=600, draw movenumber=30 movecount=5 score=5, time forfeit disabled (NO CERES losses on time)
Comment:Lc0 binary still on top here.
Comment: I put a time buffer of 9999ms on cutechess to avoid any Ceres time losses. I used @antpocas build and @masterkni666 lc0.dll files and uncompressed network. Bench results are here: https://discord.com/channels/425419482568196106/539960268982059008/794651553637072967
I can't say much on the games as I haven't check them yet. I'll post in #test-results when the 1000 games match will be over.
Comment: Sometime Ceres reports very low depth (but the node count is normal).
Games: https://gofile.io/d/YMxRfn
# PLAYER : RATING ERROR PLAYED (%) CFS W D L D(%)
1 stockfish-201225-18T : 0.0 ---- 645 51.63 55 75 516 54 80.00
2 lc0-263-66740 : -1.1 17.0 320 49.84 96 32 255 33 79.69
3 CERES-66740 : -22.0 16.3 325 46.92 --- 22 261 42 80.31

White advantage = 39.95 +/- 6.20
Draw rate (equal opponents) = 83.31 % +/- 1.71

Engine sec/move Depth MIDG EARLY ENDG LATE
CERES-66740 3.20 6.52 8.33 | 7.85 | 6.57 | 6.04
stockfish-201225-18T 3.25 41.52 34.95 | 36.03 | 44.62 | 56.24

Engine sec/move Depth MIDG EARLY ENDG LATE
lc0-263-66740 3.35 12.44 14.41 | 14.46 | 10.81 | 7.88
stockfish-201225-18T 3.31 40.64 34.93 | 33.85 | 43.51 | 57.83

https://gofile.io/d/YMxRfn

mehmet1921 · Post by **mehmet1921** » 03/01/2021, 5:10

Lc0 and Ceres still have significant development potential. Although I had an experience with Lc0 in last year. I managed to get good results like 10-20 elo with Lc0 Kayra settings against Stockfish. I had similar good scores not only in the tests I made myself but also in some independent tests. In the FGRL rating list, when Lc0 Kayra4 settings received 4-5 points lower than the Lc0 default settings, some people rightly wondered the reason of this difference. However, there was no difference. In this rating list, Lc0 Kayra had a good performance (+15 elo) after 500 games compared to the Lc0 default against Stockfish versions. The only difference was that lc0 Kayra4 got lower percentage wins against weaker engines due to being less aggressive. I actually developed a Lc0 chess engine version of more aggressive settings. But I didn't send it to be tested to Andreas Strangmüller Because this version will most likely get a better rating due to its more aggressive nature, but its performance against the Stockfish would be lower. Test results may include differences according to the personal perspective.

It is clear that Lc0/Ceres etc. also has serious potential. GPU-based chess engines will steer the future. Nvidia RTX 3080 gpu has 8714 and Nvidia RTX 3090 gpu has 10496 cuda cores. As graphics cards get better and cheaper, gpu based engines will be more dominant.

Nemesis · Post by **Nemesis** » 03/01/2021, 15:38

I had a few minutes this morning so ran Ceres in the Shredder Classic GUI using the FEN/EPD below - had the following evaluation after approx 5 minutes :

However Arena gave the following evaluation :

On my 3950X - Shredder first then Arena.
After putting my eyeballs back into their sockets I reran the position - same result.
I'll continue using my brain.

6k1/ppr5/3bp1pp/Pb1p1p2/3P3P/3NP1P1/1Pq1BPK1/2R1Q3 b - - 6 35

xrf03 [xrf04

kramnik · Post by **kramnik** » 03/01/2021, 20:21

janus wrote:

I had a few minutes this morning so ran Ceres in the Shredder Classic GUI using the FEN/EPD below - had the following evaluation after approx 5 minutes :

However Arena gave the following evaluation :

On my 3950X - Shredder first then Arena.
After putting my eyeballs back into their sockets I reran the position - same result.
I'll continue using my brain.

kramnik · Post by **kramnik** » 03/01/2021, 20:30

I guess that was a mere fluctuation; I've checked that postion and it's 0.10~0.00 at depth 41.

Outskirts CheSS ForuM

Outskirts CheSS ForuM