Lc0 v0.17.0-rc1 egt test

[George Tsavdaris]

Could you specify which positions are these?

I'm afraid i cannot do it since i don't remember now but perhaps by comparison you can do it. These are my 400 position that differ only slightly from yours:
https://drive.google.com/file/d/1HXdKeb ... 0Fkob/view

So it needs more nodes to find more best moves. In your egt test of 350/400, 50 were still missed, so perhaps there is still something to be improved in this area.

Well Leela lacks DTZ implementation so maybe with this it will improve even more.

Is this Lc0 v0.17.0 rc1 or rc2?

I have run it with V17 rc1.

I run another test with that new test file, thread is 1 and Lc0 on rc2 and new net id.

It did better with rc2 and now there is an implementation that uses DTZ i think. Not yet released officially.
How do you do your EPD tests? Arena or cutechess?

[dkappe]

Thought I'd add two data points:

on a GTX 1070, we have the following performance, without TB:

Ender 6, Sergei.S.Markoff:Tablebase.Test - 5 sec, 2 threads
Success rate: 85.53% (337/394)
T947, Sergei.S.Markoff:Tablebase.Test - 5 sec, 2 threads
Success rate: 75.38% (297/394)

Note that the Leela nets do poorly on epd's as they have no history.

Ender is a 64x6 net trained on adversarial play with sf9tb. (https://github.com/dkappe/leela-chess-w ... ndgame-Net)

[dkappe]

With @dje's no history patch and no pruning, Ender 27 without TB trails only SF on a GTX 1070.

Code: Select all

Ender 27, Sergei.S.Markoff:Tablebase.Test - 5 sec, 2 threads, no prune
Success rate: 92.13% (363/394)

https://github.com/dkappe/leela-chess-w ... ndgame-Net