Engine Tournament

[q30]

DarkNet-MoGo4: 3-1
DarkNet-Pachi: 0-4

DarkNet start up parameters:

Code: Select all

 go engine cfg/engine.cfg go.weights

Details

So DarkNet isn't ready for use (weak engine, no documentation, doesn't know nor time control commands, nor resigning, nor final score)...

The rating of strongest programs with GTP support without using GPU by real sparring is next:

League A
1. Leela

League B
2. Hiratuka
3. Ray
4. Pachi
5. DarkNet
6. MoGo

Go engines start up parameters.

[Vargo]

I'm not sure if it's the right place to post this. If not, please tell me.

I've used Sabaki for matches (GTP) between AQ 2.0.3, Leela0110GTP_OpenCL, Leela0110GTP, and GnuGo 3.8

My computer : Windows 10 i7-6700K (4gHz) 32GB GTX1080
Leela benchmark : 200000 games in 9.27s. 21574 g/s (2696 g/s per thread)

GnuGo level 10, AQ and Leela 30" per move (time_settings 0 30 1)
no pondering


Results :

9 stones handicap : Leela0110GTP_OpenCL (W) beats GnuGo (B) 5-0 (3 times by resignation, and twice by 40-50 points)

Even games : AQ beats Leela0110GTP_OpenCL 5-0

2 stones handicap : AQ (W) beats Leela0110GTP_OpenCL (B) 4-1 (always by resignation)

AQ is very strong... It seems to confirm these results :

1. AQ 2.0.3 12/16
2. Leela 0.11.0 Beta 11 4/16

2 stones handicap : Leela0110GTP_OpenCL (W) beats Leela0110GTP (B) 4-1 (always by resignation)

Leela "GPU accelerated" seems to be 2 stones stronger than the standard version...


I'm very grateful to the authors of these programs, I enjoy them immensely.
I have lots of questions about the various commands (time settings, number of threads, etc. etc) but I don't think it's the right place to ask them.

I'll keep experimenting, with other handicaps and time settings

[pookpooi]

Who is stronger between AQ, Rayon, and Leela? Can we rank them in 1-2-3 order?

[q30]

I'll keep experimenting, with other handicaps and time settings

You can try with pondering, half of Your PC CPU (I don't know, how it may be with GPU) cores to each engine and 2' for move without handicap (it will be close to real pro games)...

[q30]

Who is stronger between AQ, Rayon, and Leela? Can we rank them in 1-2-3 order?

If Leela without OpenCL - I think, it will be 3. But it beats Ray...

[as0770]

You can try with pondering, half of Your PC CPU (I don't know, how it may be with GPU) cores to each engine and 2' for move without handicap (it will be close to real pro games)...

You can't play fair ponder on matches on one GPU.

[Shoreline]

Who is stronger between AQ, Rayon, and Leela? Can we rank them in 1-2-3 order?

Don't know anything about AQ. The DGS ranking of Leela and Rayon bots have been consistently showing Rayon with 2 stone stronger.

Leela v0.10.o , 4 Dan
http://www.dragongoserver.net/userinfo. ... LeelaBot19

Rayon version 8.0.1 (RN-3.9), 6 Dan
http://www.dragongoserver.net/userinfo. ... r=rayonbot

[Vargo]

Leela v0.10.o , 4 Dan

Leela 0.10 is weaker than Leela 0.11, which is weaker than Leela 0.11 openCL, which is weaker than AQ.

On my computer, AQ 2.0.3.1 (W) wins each time against Rn-4.20.4 (B) with 2 stones handicap

In CGOS (computer go server) http://www.yss-aya.com/cgos/19x19/bayes.html
AQ is 3877 Elo, it's very high. On the same site, for example, Gnugo-level 11 is ranked 517, with 1825 Elo, and
Ray-4.20, running on 8 CPU, is 3743 Elo
Still in the same site, Leela is around 3700 Elo with 16 threads and GPU.
But it's heavily dependent on the hardware, it's hard to make a fair comparison, because we usually don't know the exact specifications of the computers used for these games, hence the interest of the results obtained by as0770 in his engine tournament, same conditions and same hardware used for every engine (I suppose)

[splee99]

Leela 0.11 gave Hiratuka 10.37B 4 handicaps and still won.

[pookpooi]

So in conclusion
Latest AQ >> latest Rayon
Latest Rayon > Latest Leela
Latest Leela >>>> Latest Hiratuka

[as0770]

League B updated with Leela Zero and the weights file provided by GCP. It will be interesting to see how it can improve with better weigths files.

Interesting that it lost 4 of 4 vs. Fuego and even 2 of 4 vs. Orego. On the other hand it was able to win 2 of 4 against Ray.

New entry in League D is GoTraxx.

Leela vs. AQ

Code: Select all

    1. AQ 2.0.3                     12/16
    2. Leela 0.11.0 Beta 11          4/16

Configuration:

2h/game, pondering off, 4 threads, 2GB on 4 x Intel® Core™ i5-4210H CPU @ 2.90GHz, 8 GiB Ram and GeForce 840M/PCIe/SSE2

AQ: AQ

Code: Select all

aq_config.txt:
-main　time[sec] =7200
-time controll =true
-japanese rule =true

Leela: leela_gtp_opencl –gtp –threads 4

twogtp: gogui-twogtp -auto -white "leela_gtp_opencl --gtp --threads 4" -black "AQ" -games 4 -size 19 -time 120 -verbose -sgffile xxxx

League A:

Code: Select all

    1. Leela 0.10.0                 22/24
    2. Rayon 4.6.0                  19/24
    3. Oakfoam 0.2.1 NG-06          18/24
    4. Hiratuka 10.37B (CPU)         9/24
    5. DarkForest v2 MCTS 1.0        7/24
    6  DarkGo 1.0                    5/24
    7. Pachi DCNN 11.99              4/24

Configuration:

2h/game, pondering off, 4 threads, 2GB on 4 x Intel® Core™ i5-4210H CPU @ 2.90GHz, 8 GiB Ram and GeForce 840M/PCIe/SSE2

DarkForest: taskset -c 0 bash cnn_evaluator.sh 1 /data/local/go and taskset -c 0 th cnnPlayerMCTSV2.lua --num_gpu 1 --num_tree_thread 1 --rollout 2500 --win_rate_thres 0.1
DarkGo: darknet go engine cfg/go.test.cfg go.weights
Hiratuka: Hiratuka-19×19.exe -po 175000
Leela: leela_gtp_opencl –gtp –threads 4 –noponder
Oakfoam: oakfoam -c nicego-cnn-06.gtp

Code: Select all

nicego-cnn-06.gtp:
param playouts_per_move_max 40000
param thread_count 4

Pachi: pachidcnn -f pachibook.dat threads=4,max_tree_size=2048,pondering=0
Rayon: rayon --thread 4 --no-debug

twogtp: TWOGTP=”gogui-twogtp -black \”$BLACK\” -white \”$WHITE\” -games 2 -size 19 -time 120 -sgffile xxxx”
gogui -size 19 -program “$TWOGTP” -computer-both -auto

League B:

Code: Select all

    1. Pachi DCNN 11.99             29/32
    2. Ray 9.0.1                    27/32
    3. MoGo 4.86                    21/32
    4. deltaGo 1.0.0                17/32
    5. Fuego 1.1                    17/32
    6. Leela Zero 0.1               15/17   
    7. Michi C-2 1.4.2               9/32
    8. Orego 7.08                    7/32
    9. GNU Go 3.8                    2/32

Configuration:

1h/game, pondering off, 1 thread, 1GB on 4 x Intel® Core™ i5-4210H CPU @ 2.90GHz, 8 GiB Ram and GeForce 840M/PCIe/SSE2

deltaGo: deltaGo.exe
Fuego: fuego.exe –config fuego.cfg

Code: Select all

fuego.cfg:
uct_param_search number_threads 1
uct_param_search lock_free 0
uct_max_memory 1024000000
uct_param_player reuse_subtree 1
uct_param_player ponder 0
uct_param_player early_pass 1

GnuGo: gnugo --mode gtp --level 10 --resign-allowed
Leela Zero: leelaz --gtp --w /path/to/weights.txt --noponder
Michi C-2: michi gtp

Code: Select all

ui.c:
init_large_patterns("patterns2.prob", "patterns2.spat"); // Michis pattern files renamed because they have the same name as Pachis files.

MoGo: mogo
Orego: java -jar /path/to/orego-7.08.jar threads=1 grace
Pachi: pachi -f pachibook.dat threads=1,max_tree_size=1024,pondering=0
Ray: ray --time 3600 --thread 1 --no-debug
twogtp: TWOGTP=”gogui-twogtp -black \”$BLACK\” -white \”$WHITE\” -games 2 -size 19 -time 60 -sgffile xxxx”
gogui -size 19 -program “$TWOGTP” -computer-both -auto

League C:

Code: Select all

    1. GNU Go 3.8                   24/28
    2. Hara 0.9                     18/28
    3. Dariush 3.1.5.7              16/28
    4. Indigo 2009                  15/28
    5. Matilda 1.24                 15/28
    6. Aya 6.34                     11/28
    7. Fudo Go 3.0                  11/28
    8. JrefBot 081016-2022           2/28

Configuration:

1h/game, pondering off, 1 thread, 1GB on 4 x Intel® Core™ i5-4210H CPU @ 2.90GHz, 8 GiB Ram and GeForce 840M/PCIe/SSE2

Aya: Aya.exe –mode gtp –level max
Dariush: DarGTP.exe –level 10
Fudo Go: taskset -c 0 fudo –boardsize=19 –komi=6.5
GnuGo: gnugo --mode gtp --level 10 --resign-allowed
Hara: hara
IndiGo: Indigo.exe -gtp
JrefBot: java -jar jrefgo.jar 10000
Matilda: matilda

Code: Select all

matilda.h:
#define BOARD_SIZ 19
#define DEFAULT_UCT_MEMORY 1000
#define DEFAULT_NUM_THREADS 1

twogtp: TWOGTP=”gogui-twogtp -black \”$BLACK\” -white \”$WHITE\” -games 2 -size 19 -time 60 -sgffile xxxx”
gogui -size 19 -program “$TWOGTP” -computer-both -auto

League D:

Code: Select all

    1. JrefBot 081016-2022          29/32             
    2. Iomrascálaí 0.3.2            25/32
    3. Crazy Patterns 0008-13       23/32
    4. Marcos Go 1.0                22/32
    5. AmiGo 1.8                    20/32
    6. Stop 0.9-005                 10/32
    7. CopyBot 0.1                   6/32
    8. GoTraxx 1.4.2                 6/32
    9. Brown 1.0                     3/32

Configuration:

1h/game, pondering off, 1 thread, 1GB on 4 x Intel® Core™ i5-4210H CPU @ 2.90GHz, 8 GiB Ram and GeForce 840M/PCIe/SSE2

Amigo: amigogtp
Brown: brown.exe
Copybot: python /path/to/__main__.py
CrazyPatterns: CrazyPatterns.exe
GoTraxx: GoTraxx.exe
Iomrascálaí: taskset -c 0,1 iomrascalai
JrefBot: java -jar jrefgo.jar 10000
Marcos Go: marcos_go --patterns /path/to/patterns.txt --cycles_mcts 10000 --threads_mcts 1
Stop: /usr/bin/java -ea -jar /path/to/stop-09-005.jar --mode gtp

twogtp: TWOGTP=”gogui-twogtp -black \”$BLACK\” -white \”$WHITE\” -games 2 -size 19 -time 60 -sgffile xxxx”
gogui -size 19 -program “$TWOGTP” -computer-both -auto

Links:

Amigo: https://sourceforge.net/projects/amigogtp/
AQ: https://github.com/ymgaq/AQ
Aya: http://www.yss-aya.com/
Brown: http://ricoh51.free.fr/go/engineeng.htm
Copybot: https://github.com/sirtango/ICopyMoves
CrazyPatterns: https://www.remi-coulom.fr/Amsterdam2007/
Dariush: http://ricoh51.free.fr/go/engineeng.htm
DarkForest: https://github.com/facebookresearch/darkforestGo
DarkGo: https://pjreddie.com/darknet/darkgo-go-in-darknet/
deltaGo: http://home.q00.itscom.net/otsuki/delta.html
Fudo Go: http://www.geocities.jp/hideki_katoh/
Fuego: http://fuego.sourceforge.net/
GnuGo: https://www.gnu.org/software/gnugo/devel.html
GoTraxx http://gotraxx.codeplex.com/
Hara: https://github.com/antoniogarro/Hara
Hiratuka: Non GPU version (10.37B): http://www.vector.co.jp/download/file/winnt/game/fh673259.html / GPU version (10.38B): http://www.vector.co.jp/download/file/winnt/game/fh688349.html
IndiGo: http://www.math-info.univ-paris5.fr/~bouzy/INDIGO.html
Iomrascálaí: https://github.com/ujh/iomrascalai
JrefBot: http://ricoh51.free.fr/go/engineeng.htm
Leela: https://sjeng.org/leela.html
Leela Zero: https://github.com/gcp/leela-zero
Marcos Go: https://github.com/MarcosPividori/Go-player
Matilda: https://github.com/gonmf/matilda
Michi C-2 https://github.com/db3108/michi-c2
MoGo: https://lifein19x19.com/forum/viewtopic.php?p=211091#p211091
Oakfoam: https://bitbucket.org/dsmic/oakfoam
Orego: https://sites.google.com/a/lclark.edu/drake/research/orego
Pachi: http://pachi.or.cz/
Rayon: https://github.com/zakki/Ray
Ray: https://github.com/kobanium/Ray
Stop: https://www.vanheusden.com/stop/

Best,
Alex

[q30]

But it's heavily dependent on the hardware, it's hard to make a fair comparison, because we usually don't know the exact specifications of the computers used for these games, hence the interest of the results obtained by as0770 in his engine tournament, same conditions and same hardware used for every engine (I suppose)

Quite right! Except the problems of same conditions while using GPU and of obtained with this synthetic conditions results equivalenty to results, obtained with real games conditions...

[q30]

I have tested Leela 0.11.0 (details). The rating of engines with GTP support and without using GPU by sparrings in real Go games conditions remains the same:

League A
1. Leela

League B
2. Hiratuka
3. Ray
4. Pachi
5. DarkNet
6. MoGo

Go engines start up parameters.

[q30]

Darknet from 08.11.17 doesn't play - passes. So this engine will not be in new ratings...
Oakfoam remains very weak engine (weaker, then MoGo: details).
The rating of Go engines without using GPU is next:

League A
1. Leela

League B
2. Hiratuka
3. Ray
4. Pachi
5. MoGo

Go engines start up parameters.

[as0770]

Quite right! Except the problems of same conditions while using GPU and of obtained with this synthetic conditions results equivalenty to results, obtained with real games conditions...

Playing engine matches with ponder=off on one GPU is no problem. There might be a slight difference because of the GPU memory usage but this is not measurable for me.

Playing matches on different machines can always lead to different results. We can rate engines just for specific conditions. Budget- or High-End-PC. With one GPU, with several GPUs or without GPU, With one CPU or several CPUs. With timecontrol, or fixed number of playouts. The difference between ponder on and ponder off matches is nearly irrelevant compared to all the other influences and for sure interfered by the statistical fluctuation when rating the engines with less than 100 games.