AntiFish 1.0 (Mark 289)
1,080 views
Skip to first unread message
Cscuile
Feb 10, 2019, 7:33:33 AM2/10/19
to LCZero
AntiFish 1.0 (Mark 289) Download: https://github.com/Cscuile/AntiFish/releases/tag/Mk289
First off, I'd like to thank javiros and nameBrandon with the help in testing for both AF and the UTI Project (Universal Testing Initiative). While I am still unsure of how this network scales at higher TCs, this is the strongest network we currently have found. AntiFish 1.0 (Mark 289) has 20,500 Steps, 2 LRs (0.002 and 0.00002) a Policy Focus, and is trained off of 233,620 games.
First off, I'd like to thank javiros and nameBrandon with the help in testing for both AF and the UTI Project (Universal Testing Initiative). While I am still unsure of how this network scales at higher TCs, this is the strongest network we currently have found. AntiFish 1.0 (Mark 289) has 20,500 Steps, 2 LRs (0.002 and 0.00002) a Policy Focus, and is trained off of 233,620 games.
Since I am now using 2 LRs, far more games are now required which means that progress will slow even further. The networks leading up to AntiFish 1.0 have not been saturated in training because of this as we would risk overfitting if we used more steps.
To Help with the AntiFish Project Please Visit:
AntiFish Contribution Channel Invite Link: https://discord.gg/chtSb2R
Thank you to everyone who contributed to this project!
AntiFish Contributors:
Fulgore
ASilver
Dkappe
joel
Additional Pylon
hv
PaleMale53
Matuiss2
Edosan
❤ u
Corzlyn
AlejoCarlsen
Sonicast
hx
Charis
Bladewood
coolchess123
Plutie
Filipe
javiros
nameBrandon
If I had missed anyone's name, I apologize. Please tell me and I will gladly add you name here.
Ingo Weidner
Feb 10, 2019, 8:14:32 AM2/10/19
to LCZero
Like with previous versions of AntiFish and some other networks (including 32930) i just added AF 289 to my current testing gauntlet against Stockfish development version (version from February 3rd) at 15min+5s TC.
In the games played so far AF 275 was equally good as 32930 so maybe AF 289 could finally surpass it.
Saleh Aydin
Feb 10, 2019, 9:33:03 AM2/10/19
to LCZero
Will Antifish play at TCEC in season 15?
Cscuile
Feb 10, 2019, 12:26:48 PM2/10/19
to LCZero
Thank you so much for all your tests!
Cscuile
Feb 10, 2019, 12:27:56 PM2/10/19
to LCZero
Unfortunately TCEC has a very strict submission policy, they would not even allow DeusX in which is an engine that completely deviates from Lc0's weights. So the chances of TCEC allowing AF to play is quite low.
garrykli...@gmail.com
Feb 10, 2019, 8:57:32 PM2/10/19
to LCZero
Most were EXTREMELY satisfied with the decisions to exclude fakeX
smashu2
Feb 10, 2019, 10:33:54 PM2/10/19
to LCZero
It does not matter if you cant participate in TCE because if you can create a stronger net from the best net 30 that plateau you win since you show you can still increase the strength of the net once it plateau by using non zero method.
On Sunday, February 10, 2019 at 12:27:56 PM UTC-5, Cscuile wrote:
Aslam
Feb 11, 2019, 1:20:36 AM2/11/19
to LCZero
Is this AntiFish based off of 32425 or 32930?
Ingo Weidner
Feb 11, 2019, 3:47:49 AM2/11/19
to LCZero
AFAIK AntiFish until 228 was based on 32425 and newer ones up to 289 are based on 32930.
Message has been deleted
Lothar Jung
Feb 12, 2019, 3:53:30 AM2/12/19
to LCZero
How does Antifish behave in Endgames?
Reply all
Reply to author
Forward
0 new messages