I've been actively developing Invictus again after a 1.5 year hiatus. I managed to fix some bugs and introduce some better prunings which resulted to a big jump from r305 to r340. The eval is still the same. At the current stage though I think that the eval has reached it's maximun strength. I have used lichess-quiet.epd which I have downloaded from this forum (not sure who to credit with it) to tune the values.
Before I proceed into implementing my own NNUE architecture and version I would like to improve my hand crafted eval first as I will use it later to train my own NNUE net.
Do we have an available set of training positions somewhere that can be downloaded? Preferably in an "epd | game result" format.
Training positions
Moderators: hgm, Rebel, chrisw
-
- Posts: 803
- Joined: Mon Jul 17, 2006 5:53 am
- Full name: Edsel Apostol
-
- Posts: 1871
- Joined: Sat Nov 25, 2017 2:28 pm
- Location: France
Re: Training positions
A good input is indeed the lichess-quiet set on this forum or simply the texel ones for a start. Best would be your own of course!
-
- Posts: 1871
- Joined: Sat Nov 25, 2017 2:28 pm
- Location: France
Re: Training positions
Sorry, not the Texel ones. I don't know if they are available, I meant the zurichess ones.
But why not just score your ones from previous games played.
Once I also give a try to CCRL game database.
-
- Posts: 4833
- Joined: Sun Aug 10, 2008 3:15 pm
- Location: Philippines
Re: Training positions
I think the strength of Invictus is enough to generate a good training positions for NN.Edsel Apostol wrote: ↑Thu May 27, 2021 7:19 pm Before I proceed into implementing my own NNUE architecture and version I would like to improve my hand crafted eval first as I will use it later to train my own NNUE net.
Depth 5 test:
Code: Select all
Rank Name Elo +/- Games Score Draw
0 Invictus r340 d5 137 28 553 68.7% 15.6%
1 sf13 d5 nnue_yes -74 38 277 39.5% 18.4%
2 sf13 d5 nnue_no -210 45 276 23.0% 12.7%
The real challenge is what positions should be in the training positions. You can devise a wide range of randomization, tactical or not tactical, or a little bit of tactical positions, percentage of related vs unrelated positions, wider range of unique positions. Or something such that the resulting NN can infer with greater performance.
Training Hyperparameters is generally not an issue you just need more computing power to get optimal values. But finding the right positions involves a lot of trials.
Once you have an NN you can also use it to generate training positions for the new NN.