Is there a real advantage of using HalfKA over HalfKP?
Aren't the additional pieces for the kings redundant?
The test of Stockfish (https://github.com/official-stockfish/S ... abe82e0838) is not meaningful because it changes also the network size.
Has someone already tested the two different architectures? With what results?
Thank you.
Neural network architectures HalfKP vs HalfKA
Moderator: Ras
-
Fabio Gobbato
- Posts: 219
- Joined: Fri Apr 11, 2014 10:45 am
- Full name: Fabio Gobbato
-
AndrewGrant
- Posts: 1957
- Joined: Tue Apr 19, 2016 6:08 am
- Location: U.S.A
- Full name: Andrew Grant
Re: Neural network architectures HalfKP vs HalfKA
It is not redundant. If you think of KP as "King x Pawn, King x Knight, ... King x Queen", KA adds the additional "King x King", which is a set of inputs based on the King positions relative to each other. You could also do the King x Itself, which adds a more PSQT-like input.
I have not had success with KA over KP. It is near equal to slightly worse for me, and ultimately takes longer to train, so there is little to no value in it for me.
I have not had success with KA over KP. It is near equal to slightly worse for me, and ultimately takes longer to train, so there is little to no value in it for me.