Beyond Leela Zero ( Leela will get part of Stockfish blood)

[Nay Lin Tun]

COPY FROM Jhorthos

STARTING WITH J92-340, training will include a small percentage of value repair games, which use SF12 (1M nodes) to identify poor positional J92 evaluations (no search). This is clearly non-zero, so if community sentimentis to continue zero-only nets in competition, J92-330 is effectively the last one from this series. Decided to make the Value Repair a new training series J94 - coming soon

Thus being said, stockfish evaluation will be included in Leela evaluation, repairing many weakness in Leela's evaluation for TCEC season 20.

Actually, Leela is beyond zero since last year , with addition of Tablebase adjudication, addition of MLH ( knowledge about how many moves left in 50 moves rules) etc. And now with addition of Stockfish evaluation, it is well beyound zero ( Leela One)

[AdminX]

I saw this and had the same thoughts when I read it. 'Leela One' would be a great rebranding since it is really no longer zero.

[jmartus]

Sounds like cheating but I'm not an expert on that.

[Ozymandias]

Note - for practical hardware reasons I use SF12 to find bad positional evaluations, but there is every reason to expect that doing the same with Leela after search will be nearly identical in outcome.

[AdminX]

Value Repair has started.

J94 384x30b

J94-40

[Ozymandias]

Yes, he changed the nomenclature.

I've already seen a test at Discord and it's on par with 330.

[dkappe]

Sounds like cheating but I'm not an expert on that.

So this is essentially an outgrowth of something I did with Ender a few years ago. I called them “dodgy positions,” but “value repair” seems like a better name. Stockfish is only used to identify problematic positions (where the sf eval is converted to a [-1,1] range and abs(sf eval - nn eval) >= 0.5), but leela then trains on these positions. No sf data is folded into the leela training. For this purpose you could use just about any 3000+ ab engine, but doing a 100k node search on sf is speedier.

The story of Ender can be found here: https://github.com/dkappe/leela-chess-w ... ndgame-Net

The effect of mixing in a higher percentage of these positions was dramatic, almost like chemotherapy for the net. The percentage of problematic positions for Ender per 100k lichess 18 man positions went down from 12% to just above 6% over the course of maybe 400k games.

I don’t know if this will add a ton of elo, but value repair is the first experiment on a full net with a large number of hw resources behind it.

I should add that oscardssmith had the original idea of training on 6 man positions where the net disagreed with an egtb.

[Nay Lin Tun]

Value Repair has started.

J94 384x30b

J94-40

He has plan to do policy repair as well.

Previous myths about "Neural Networks are blackbox" is over. He is going to reapir NN.

This sounds like. " Genetic Engineering ".

[Nay Lin Tun]

Sounds like cheating but I'm not an expert on that.

This sounds like. " Genetic Engineering " but I am not an expert on that.