Hier der Beitrag auf Discord:
Why is there such confusion?
There are two types of parameter tuning:
1) What Kiudee method does via playouts
2) Optimising the parameters for highest speed (Which is what my "Tuner program" does.)
Example: Tilps suggestion to add threads=2 to Demux backend opts, produces an extra 10000 nps.
Currently being tested are the following parameter combinations:
```{
"seconds_per_move": 10,
"options": {
"backend-opts": ["backend=cuda-fp16,multi_stream=true,threads=2,(gpu=0),(gpu=1),(gpu=2),(gpu=3)"
],
"multi-gather": ["true"
],
"backend": ["demux"
],
"threads": [2,3
],
"nncache": [50000000,20000000,30000000,40000000
],
"minibatch-size": [816,408
],
"max-prefetch": [37,160,32,48,64
],
"cpuct": [2.0179,1.745,0.95
],
"cpuct-base": [50000,38739
],
"cpuct-factor": [3.8542,3.894
],
"fpu-value": [0.4089,0.33,0.25
],
"temperature": [1.4229,1.359
],
"smart-pruning-minimum-batches": [300,0
],
"max-out-of-order-evals-factor": [2.4,2,3
],
"smart-pruning-factor": [3.6141,1.33
],
"time-manager": ["alphazero(alphazero-time-pct=23.06290)",legacy
],
"verbose-move-stats": ["true","false"
],
"ponder": ["false"
],
"move-overhead": [0
]
}
}```
And there are 9216 configuration combinations of the above.
Lothar