Lc0: A Chess Engine that Learns from Self-Play
Running out of memory Unlike traditional A/B chess engines, Lc0 has to keep entire search tree in memory. That means that as the search goes, memory consumption increases. Lc0 requires roughly 250-350 bytes of memory per node, that means that typical RAM sizes can only handle so many nodes:
lc0 chess engine download
On Windows 10 I have installed the latest versions of Arena (v3.5.1) and the Lc0 engine (v0.23.1). However I have only found one weights.pb file that it will accept. The one in the link in this Q&A. But this seems rather weak, Stockfish beats it every time and it seems to make silly mistakes, particularly in the end game.
The version of SF I'm using is that which comes bundled with Arena, Stockfish_9_x64. Both engines are configured to run with 12 "common max CPU cores" under the UCI tab - Options for all UCI Engines. (I've got a Ryzen 8+8 core CPU) They are paired together in the default Engines > Tournament with 10 rounds of 5 minute games.
But the real problem isn't that SF beats Lc0 10-0, it's before they even get to play. Lc0 rejects all but one weights file when I am in (F11) Engine Management > Details, change the -w configuration on the command line and watch it throw the error in the debug window (F4) when I click on "Start this engine right now!". If I could solve this, I could try some different weights until I find one I like.
lc0 chess engine download windows
lc0 chess engine download mac
lc0 chess engine download android
lc0 chess engine download linux
lc0 chess engine download github
lc0 chess engine download cuda
lc0 chess engine download cudnn
lc0 chess engine download onnx-dml
lc0 chess engine download dnnl
lc0 chess engine download openblas
lc0 chess engine download homebrew
lc0 chess engine download latest version
lc0 chess engine download network file
lc0 chess engine download training client
lc0 chess engine download source code
lc0 chess engine setup guide
lc0 chess engine best gui
lc0 chess engine performance test
lc0 chess engine rating list
lc0 chess engine configuration options
lc0 chess engine vs stockfish
lc0 chess engine vs alphazero
lc0 chess engine vs komodo
lc0 chess engine vs houdini
lc0 chess engine vs leela zero go
lc0 neural network based chess engine
lc0 open source neural network chess engine
lc0 deep learning neural network chess engine
lc0 self-learning neural network chess engine
lc0 reinforcement learning neural network chess engine
lc0 monte carlo tree search neural network chess engine
leela chess zero official website
leela chess zero play online
leela chess zero watch live games
leela chess zero contribute to training
leela chess zero development blog
leela chess zero project history
leela chess zero project goals
leela chess zero project team
leela chess zero project supporters
leela chess zero uci protocol implementation
leela chess zero uci options explained
leela chess zero uci commands reference
leela chess zero uci compatible gui list
leela chess zero uci compatible engines list
Since you didn't specify some things, I'm just going to give instructions from scratch. Since you didn't specify which version of lc0 you downloaded, make sure that it's 23.2 (from ). If you only have a cpu you want -v0.23.2-windows-cpu-dnnl.zip. For cpu only, the best net is no longer LD2, although it's still pretty strong. A current list of best nets of different sizes is at -Nets-for-Lc0. At time of writing, current best 10b net is probably 58599 ( _network?sha=2b482ff696e3a4032be1b7477567661527a9fac4166591167bf0a71183f9a669). Based on what you were reporting, I'm guessing you probably had an old version of lc0 that is incompatible with newer nets. It will likely still be a few hundred elo worse than SF without a gpu, but it should at least put up a fight.
Filip Höfer chess software has a built-in chess engine (i.e., the artificial intelligence providing computer's and coach's moves). Chess 2020 and newer products also have the capability to switch from this built-in engine to an external one. For this purpose, the software implements the UCI standard. In practice, the user goes to the Level menu, selects Computer or Coach and then UCI engine. In the UCI engine dialog, it is necessary to point to an executable file (*.exe) of the desired external engine. Here are some engines that have been tested with Filip Höfer chess software and can be freely downloaded:
In the UCI engine dialog, it is possible to configure Computation depth and Computation time. By default, an UCI engine is on a mission to compute the perfect move. For most positions, this would take an extreme amount of time. Therefore, there are options to limit the engine by time, by depth, or both. As soon as one of the set limits is reached, the engine stops and returns the best move it has been able to compute until that point.
The time in milliseconds actually also determines the depth. But for a fixed time, the depth may vary given the speed of the machine and the complexity of the chess position. On a medium-performance laptop, it takes about 20 seconds to get to depth 22 from the initial position. The depth-time relationship is actually exponential, so even one second should be enough to get very good results, perhaps around depth 10 or so.
Some engines are provided in multiple versions to accommodate for different hardware. Typically, these versions are sorted from the least compatible (but fastest) to the most compatible (but slowest). You can go down the list and use the first binary that works.
OK my friend, after taking a few mins to look around I found an alternative download to the weight file and updated the PKGBUILD. Use the following and you should be right. I was able to build lc0 without a problem 2b4c41e320