TalkChess.com

If you want some smaller leela nets that run better on PC, you can check out some of the 1CPU tests running under CCRL 40/40 conditions. viewtopic.php?f=6&t=69633

Right now I’m trying to find the optimal size. Once found I will distill the best of t30, Ender, and whatever else doesn’t get out of the way fast enough.

There are two parts to leela or other neural network based engines: search and evaluation. Leela (and many go programs) use a form of search called mcts, or Monte Carlo tree search (See UCT). It’s fundamentally different from the minimax search in an engine like stockfish. It uses a strategy (in it’...

Someone made a patch for lc0 v19.1 that switched nets at a certain piece count and changed cpuct, policy softmax and fpur at the same time. I don't think it's been ported to v20.x Also, just released 11258-112x9-se. Best one yet. # PLAYER : RATING ERROR POINTS PLAYED (%) CFS(%) W D L D(%) 1 ethereal...

Note, Ender only plays endgames. It’s opening and middlegame play is laughable.

If you want a small net that plays a better endgame, try Ender. https://github.com/dkappe/leela-chess-w ... ndgame-Net

I’m going to distill it down from 128x10 to 112x9 when my big distillation is done.

It trained with egtb, so it’s designed to play 16p or less with 6p tb.

64x6 or 64x8? The 96x8-se is already released. 112x9-se is at almost 140k of 200k steps.

Next is 144x11-se.

I’m in the process of finding the optimal dimensions for a cpu-only network. The approach I am taking is to distill a 20 block network into a smaller size using the larger network to train the smaller one. My network of choice is 11258. I am using se–film networks, Which are somewhat stronger than t...

Under the parameters tab, you can select several other nets, including 11248.

Ender is a 128x10 Leela compatible net trained to play up to 16 piece endgames. It is so not zero. Major developments. The training now distills problem positions, i.e. ones where the net misevaluates value. Those positions are then overrepresented in the training. It is now the stronger Ender net y...