-
Notifications
You must be signed in to change notification settings - Fork 8
Niner Opening Middlegame Net
dkappe edited this page Oct 19, 2018
·
6 revisions
Starting my building of the 64x6 opening/middlegame net. I’m taking 9149 as the starting point. I was surprised at how weak it was. I ended up at sf9 150k nodes vs 9149 30k nodes as a benchmark. Noomen 3 man, played twice with colors reversed. Both with 6 man egtb.
Score of stockfishTB vs Niner0: 130 - 89 - 181 [0.551] 400
Elo difference: 35.74 +/- 25.22
Before self play, I decided to train 9149 against the last 2M+ self-play games of the test10 20b series. But the result was pretty bad. I decided to mix 20k test10 chunks with 20k self-play games from niner.
bender.egtb selfplay --training --games=20000 \
-w ~/data/chess/niner/niner-latest.txt.gz \
--visits=800 --cpuct=3.2 \
--syzygy-paths=/home/dkappe/data/chess/egtb --parallelism=8 \
--temperature=1 --tempdecay-moves=15 \
--resign-playthrough=10 --resign-percentage=10
The last known learning rate from the test 9 series was 0.001. I decided to go with 0.0001.
My new (old) blog is at lczero.libertymedia.io