Hello,
Does anyone have experience training a large lc0-style net (i.e. residual blocks with value and policy)? If so, I would like to know if it is possible to do an RL-based approach? If not, where can I get high quality training data? How long does the training process take for both approaches?
lc0-style net
Moderator: Ras
-
AAce3
- Posts: 80
- Joined: Fri Jul 29, 2022 1:30 am
- Full name: Aaron Li
-
smatovic
- Posts: 3224
- Joined: Wed Mar 10, 2010 10:18 pm
- Location: Hamburg, Germany
- Full name: Srdja Matovic
Re: lc0-style net
I guess this was solved already, nevertheless:
1. Lc0 -> 0 -> zero knowledge, means, like A0, it uses RL, reinforcement learning, not SL, supervised learning.
There were/are other approaches, like Deus X or Maia.
2. Lc0 and SF training data are out there in the web, f.e.:
https://github.com/glinscott/nnue-pytor ... g-datasets
https://robotmoon.com/nnue-training-data/
3. One thing is training the neural network (weights), depends on GPU and used data set++, other thing is to generate the RL/SL games, another thing is to test the neural net for Elo strength. Smaller nets saturate faster, bigger nets need more training data.
--
Srdja
1. Lc0 -> 0 -> zero knowledge, means, like A0, it uses RL, reinforcement learning, not SL, supervised learning.
There were/are other approaches, like Deus X or Maia.
2. Lc0 and SF training data are out there in the web, f.e.:
https://github.com/glinscott/nnue-pytor ... g-datasets
https://robotmoon.com/nnue-training-data/
3. One thing is training the neural network (weights), depends on GPU and used data set++, other thing is to generate the RL/SL games, another thing is to test the neural net for Elo strength. Smaller nets saturate faster, bigger nets need more training data.
--
Srdja