Getting good results from aart-nets

Discussion of anything and everything relating to chess playing software and machines.

Moderators: hgm, Rebel, chrisw

User avatar
AdminX
Posts: 6340
Joined: Mon Mar 13, 2006 2:34 pm
Location: Acworth, GA

Getting good results from aart-nets

Post by AdminX »

I am seeing better results on CPUs and Low End GPUs. Using 32930-112x9-se.pb.gz I was getting drawn matches with Wasp 3.60 using 4 threads (LC0 on my low end MX150 GPU). Using 128x10-cos-aart-swa-222000.pb.gz, LC0 is now winning these matches.
Read Me wrote:All "aart" nets here are 10x128 T50 architecture nets trained on T40 training data, with some differences in training parameters. 128x10-cos-q-franken-aart-swa is the latest training run. (ongoing) and roughly equal to cos-aart-swa-222000. it does not have WDL head to be able to train with q_ratio.

cos-aart-swa-222000.pb.gz is the net used at tcec for cpu eval with "!eval lc0" in chat.

I believe these are stronger than T50. see the attached ordo of 800 node matches done by jhorthos against 50696.

the t50swa.pb.gz is a swa net of a few of the latest ~100 nets of T50.
https://mega.nz/#F!3Q5hgACQ!OOCrMaHjRsZLIniCvjhPPw

https://mega.nz/#!6dwTwCxb!pYlo7A9xiVT7 ... NiEG51wD4Y
"Good decisions come from experience, and experience comes from bad decisions."
__________________________________________________________________
Ted Summers