TalkChess.com

Guys, I'm a layman, but how is the Stockfish neural network trained? We know that Leela's network is trained from zero, but is NNUE training supervised? How is it done? It is trained at positions evaluated by Stockfish with shallow searches. Originally by "classical Stockfish" and now by NNUE Stock...

let's suppose that the PV contains a repetition 1) e4 Nf6 2) Nf3 Ng8 3) Ng1 Nf6 4) d4 e5 .... the Pv extracted from transposition table can be eiter 1) e4 Nf6 2) Nf3 Ng8 3) Ng1 Nf6 4) Nf3 Ng8 5) Ng1 Nf6 6) Nf3 Ng8... or 1) e4 Nf6 2) d4 e5 .... because for repeated position, only one move can be sto...

I'm thinking to collect the PV by a somewhat modernized version of the triangular array method (vectors instead of arrays, because an array would need to be 256x256 elements in my case. My MAX_PLY is 255.) So I can have a PV without finding it in a hash table. I don't think you have the PV until yo...

I may be misunderstanding this, but as I understand it now, it seems "walking the PV" (and thus, principal variation search?) For me, PVS means zero-window searches of all non-first moves. This is very useful. It is fine to use the search the TT move first. You will anyway not have the PV until aft...

Type the following as root: # echo 1 > /proc/sys/vm/drop_caches Yes, this solved it. Afterwards "bin/olithink589 performs exactly equal as "./olithink589" with 2220 knps. Now this should explain the issue: What cache went probably wrong and how can this be avoided in the future? Nothing went really...

Type the following as root:
# echo 1 > /proc/sys/vm/drop_caches

Now try again.

An NNUE net based on pure PST would have far fewer inputs and weights, but would not give spectacular computational savings: the only thing you save is the complete recalculation on the occasional King move. And with 9 King positions you would have to recalculate on a King move anyway. By making th...

If you care about worst-case behaviour then all engines suck without exception, and all humans too. Indeed, and this has always worried me. No one seems to care; they sacrifice worst-case for a better average (= Elo) all the time. I think in most areas it is more fruitful to worry about the average...

The flip takes place when converting the incrementally updated accumulator to the input values for the first hidden layer. If white is to move, the white half of the accumulator gives coefficients 0-255. If black is to move, the black half of the accumulator gives coefficients 0-255. You mean the K...

And I am still confused about what the two halves are for. We are talking about white and black, but I thought originally it was stm and opponent. The latter would require quite a bit more complex handling, as basically you would have to flip the board each ply, making the inputs completely differe...