LCO Input Format

[Ben Lewis]

Hi, I understand if this is not totally relevant and no worries if you just delete my message. But I've been trying to understand how leela actually reads positions and converts them into inputs for the CNN. I know that alpha zero took a stack of previous positions matrix of 1s and 0s but that explanation only works for GO when there is only one kind of piece.

 

I've always had the goal of creating a learning algorithm that can beat me at chess and I keep coming back to it but have never been able to do it. I wanted to try do it in python . I've got an abomination of RES net but I'm not sure what to put into it or what to try and get out of it for the policy head.  I know you guys are much smarter than me and I thought any advice you could give me would be really helpful. 

[esch...@gmail.com]

Basically, it does a stack of 1s and 0s for each piece type.  So you end up with the same type of array as you would with go, but with more inputs, because you have a set of inputs for white pawns, another set of inputs for white knights, another set for black knights, etc.

[DBg]

The "silly" questions are the best. They force people to think again beyond the 0s and 1s.  back to human land... don't apologize.  I feel less lonely, I don't get to ask them, so I can get answers without having to also apologize... :)  but a 0 and 1 answer, is not really stimulating back.  I wish I was already having a toy like yours..  that would learn to at least beat me....  (but it think it needs some of its own silly time).  Good catch on the Go having only 1 piece.  but asking from 0 and 1, got you a 0 and 1 non stimulated answer.