King buckets for NN evaluation

[Fabio Gobbato]

I've seen that stockfish has 32 king buckets for the nn evaluation while berserk for example use a 16 king buckets.
16 king buckets have less weights and needs less positions for the training but what is the difference in strength of the 2 solutions? Has someone already tried both with what result?
Thank you

[jhonnold]

With horizontal mirroring, SF is using a full 64 squares for their "king bucketing" while Berserk (uses HM as well) has 32 total "king buckets". I've tested using all 64 separately and the networks are nearly equivalent, not even 1 Elo on UHO.

That being said it's been a long time since that test, so it may not be accurate anymore.

[AndrewGrant]

Regarding "32 buckets", its actually what I've called mirrored-halfka/mirrored-halfkp.

First person to test this, although Stockfish beat me to the commit:
https://github.com/AndyGrant/EtherealDe ... 3d92248cd0
The whole point is just to save your self some time in training, cut down the size of the weights.
The elo gains are basically nil. The gains in that commit are almost exclusively from some speedups.

I've not played with using more proper buckets, where you genuinely consider the king being on squares (A, B, C, D) as the same thing. You can cut the network sizes down by a lot from doing this. Perhaps even gain elo, if you pick a good scheme. For example, the white King is almost never on a7/a8/b7/b8. Perhaps joining those together would help get a better representation while training. Perhaps it would hurt you. Can't know without trying it.

[Sopel]

I checked some of my very old tests and the impact on playing strength was pretty much 0

https://docs.google.com/document/d/1gTl ... z9qmdd6mij

I don't remember if it was tried on fishtest. I had some concerns about play in endgames (where the king is in the less accurate buckets) but never investigated. Might be worth revisiting if we ever go with nets larger than now.

edit. remembered doing this for lichess https://tests.stockfishchess.org/tests/ ... ed71812a7c, though I'm 100% sure it was trained in the exact same way