Hier auf GitHub:

https://chess-tuning-tools.readthedocs.io/en/latest/index.html

Dort wird Installation und Handhabung erklärt:

In order to be able to start the tuning, first create a python environment and install chess-tuning-tools by typing:

pip install chess-tuning-tools
Furthermore, you need to have cutechess-cli in the path. The tuner will use it to run matches.

To execute the local tuner, simply run:

tune local -c tuning_config.json
Take a look at the usage instructions and the example configurations to learn how to set up the tuning_config.json file.

Hier ein einfaches Anwendungsbeispiel:

   {
       "engines": [
           {
               "command": "engine1",
               "fixed_parameters": {
                   "Threads": 2
               }
           },
           {
               "command": "engine2",
               "fixed_parameters": {
                   "Threads": 4
               }
           }
       ],
       "parameter_ranges": {
           "CPuct": "Real(0.0, 10.0)"
       },
       "engine1_tc": "1+0.1",
       "engine2_tc": "1+0.1",
       "rounds": 20,
       "opening_file": "openings.pgn"
   }

Für Lc0 erfolgt das zz-Tuning durch die Chess-Tuning-Tools von kiudee.

Hier die Links dazu:

https://github.com/kiudee/chess-tuning-tools/releases

https://chess-tuning-tools.readthedocs.io/en/latest/

Die Ergebnisse werden auf Discord auf dem Unterbereich „tune-results“ veröffentlicht:

https://discord.com/channels/425419482568196106/766262364949315614/911901584894615582

Folgende Parameter sind derzeit insbesondere davon betroffen:

                        tuned  default
CPuct                1.473    1.745
CPuctFactor       3.973    3.894
CPuctBase         45669   38739
FpuValue           0.279    0.330
              tuned (≈60Knodes)  tuned (≈10Knodes)   default
PolicyTemperature       1.295             1.206            1.359

Hier ein ausführlicher Wikipedia Artikel zum „Automated Tuning“:

https://www.chessprogramming.org/Automated_Tuning

Hier das Stockfish Tuning Kit:

https://rebel13.nl/download/stk.html

Hier die Stockfish Tuning Method:

https://www.chessprogramming.org/Stockfish%27s_Tuning_Method

Hier die Homepage der umfangreichen „Stockfish Testing Queue“:

https://tests.stockfishchess.org/tests

Hier eine wissenschaftliche Veröffentlichung zu „ Bayesian statistics approach to chess engines optimization“:

https://arxiv.org/pdf/2205.15602.pdf

Is "clippy" still concidered reliable ?
For games with ave. 1 sec/move to 15 secs/move should I use mbs=30 even though the nps is significantly slower ?```
{Lc0 --help}  --minibatch-size=1..1024
How many positions the engine tries to batch together for parallel NN
computation. Larger batches may reduce strength a bit, especially with a small
number of playouts.
[UCI: MinibatchSize  DEFAULT: 256  MIN: 1  MAX: 1024]

backendbench --clippy
/  \
|  |    __________________________________________
+  +   |                                          |
(@)(@) _| Recommended minibatch-size for this net: |
|  |  \        1s/move   (Bullet):       30       |
|| |/  |       15s/move  (Rapid):        30       |
|| ||  |       3min/move (Tournament):  226       |
|\_/|  |__________________________________________|
\___/

benchmark {1x RTX 3080 10GB , BT2 net}
===1sec===mbs30===========   ===5sec===mbs30===========
Total time (ms) : 34268      Total time (ms) : 170263
Nodes searched  : 203880     Nodes searched  : 1178770
Nodes/second    : 5949       Nodes/second    : 6923

===1sec===mbs226==========   ===5sec===mbs226==========
Total time (ms) : 35725      Total time (ms) : 171735
Nodes searched  : 254847     Nodes searched  : 1466203
Nodes/second    : 7133       Nodes/second    : 8538```

Hier das aktuelle Tuning von BT2:

**Tune of search parameters (BT2)**
**LC0 version:** v0.30.0-dev + dag-master-pr1821 (SHA: a24e47)
**LC0 options:** Network: BT2-768x15-swa-3250000, MinibatchSize=80, Backend=cuda-fp16, MoveOverheadMs=0, StrictTiming=true
**SF options:** SF-dev, Threads=6, Hash=16, "Move Overhead"=0, SyzygyProbeDepth=10, SyzygyProbeLimit=6
**Tuning ranges:** CPuct: [1.25,2.5] (extended to [0.75,2.5] after 340 iterations), FpuValue [0.0,0.75], PolicyTemperature [0.75,1.75]
**Tuning configuration:** acq function: mes, 530 iterations/26500 rounds/53000 games
**Hardware:** Ryzen 5 3600 (6x3.6GHz) + RTX 4090@300W
**Time control:** 3s/game+0.05s/move (LC0), 2.4s/game+0.04s/move (SF)
**Speed:** ≈3000 nodes/move (LC0)
**Book:** unbalanced 4-move book (80-100cp)
**Tablebases:** 6-man
**Adjudication:** 6-man TBs, -resign movecount=3 score=300, -draw movenumber=1 movecount=4 score=5
**Software:** chess-tuning-tools 0.9.3, Cutechess (restart=on)
**Optimum found:**
```
                  tuned        default
CPuct             1.287         1.745
FpuValue          0.233         0.330
PolicyTemperature 1.175         1.359``````
90.0% confidence intervals of the parameters:
Parameter          Lower bound  Upper bound
-------------------------------------------
CPuct                    1.047        1.785
FpuValue                  0.16         0.55
PolicyTemperature         1.02         1.57```