Hot answers tagged

17

You can't really objectively answer this question, but I'll share my view. One of the important factors to take into account is that Leela Chess Zero project requires tremendous computational resources to complete. The way project attains those resources is by convincing large number of people to donate part of their machine time to that project. To make ...


12

Well, it is a small sample, but assuming that there are a lot more games like these, I think it could be the following things. First, I am not sure when we first humans first decided that space was an advantage, but for as long as I have been playing, it has been a known factor. Both of these openings cede space compared to double-king-pawn openings and ...


11

Leela works by performing an exceptionally sophisticated positional evaluation at a relatively shallow search depth, whereas most chess engines work by performing a simple evaluation at as deep a search as possible. In theory this should produce a more positional style of play, and it does seem to be an effective strategy versus today's best conventional ...


10

Leela is special because it's the first strong open-source deep-learning engine. Defeating Stockfish (and also Komodo and Houdini) convincingly on all settings is a very significant milestone, practically more important than Google's AlphaZero. Google's system is inaccessible to anyone but themself. Human-understandable evaluation function has always been ...


10

Because of the enormous skill difference between these computers and humans, any kind of analysis will inevitably be post-hoc. We can tell ourselves stories about how "Stockfish should have [insert plan]," (and I'm sure some people here will) but ultimately I think that any story we could come up with would be flawed at the level of Leela/Stockfish. This isn'...


9

Generally speaking, Leela tends to have a better "intuition" and Stockfish is very good at brute force calculations. So in a structure like the French/Caro-Kann, where calculation becomes less important and strategy becomes more important, Leela will tend to do better.


8

In short, because DeepMind has TPUs. Lots and lots of TPUs. By far, the most time-consuming part of the training process is generating self-play games. DeepMind used 5000 TPUs to generate self-play games, which is a lot of processing power. Meanwhile, Lc0 crowdsources processing power from volunteer clients, which is much slower. From the DeepMind paper ...


6

Can't comment on hardware but will say something about Leela's strengths and weaknesses - If you look through the games from the TCEC superfinal, sometimes Leela is brilliant. For example in this position from game 61: [fen "r1bbnr2/pp1n1q1k/3p4/2pPp1pP/2P1PpP1/2NQ1N2/PP2BB2/2KR2R1 w - - 10 26"] Stockfish (White) evaluated this as +0.63, while Leela was ...


6

Lc0 is fully compatible with UCI. It's strength depends a ton on what hardware it's being run on. On any hardware (tested down to Raspberry Pi), it will be superhuman (over 2900 elo). With a RTX 2060 ($200ish) it should be roughly even with SF on 8 cores (although you may need to mess with settings a bit to get it there). A 2080ti puts decently (20 or so elo)...


6

This is happening because AlphaZero and Leela aren't playing against the same Stockfish. If you read the paper, AlphaZero beat Stockfish 8 (it also played a series against Stockfish 9 - same logic applies however). Both these versions of Stockfish are old, and significantly inferior to the latest version of Stockfish. AlphaZero beat Stockfish 8 by +155 -6 =...


4

Anytime a group of people get extremely hyped about something (and especially when they aggressively bash alternatives), there's a good chance this is just human tribalism at work. Leela may become extremely strong in the near future, but to say with 100% certainty this will happen now is naive. People also like change. Stockfish is a revolutionary engine (...


4

The bigger weights file corresponds to a larger Neural Network, which means more computation per node, but better evaluation per node.This is expected, and there is nothing that can make a bigger NN as fast as a smaller NN. The best solution is to use a small NN for blitz time control. I would personally recommend LD2 (available at Lc0.org/LD2)


3

This seems to be somehow LeelaFish's approach: https://github.com/killerducky/lc0/wiki/LeelaFish More about it here (and probably on more posts in the same forum): http://www.talkchess.com/forum3/viewtopic.php?t=70803&start=10


3

As you noted, w_i is not calculated by the search tree. It is simply the number of wins out of the total simulations performed at that node (if this is the method used to score the node). This scoring mechanism is also known as playout.


3

As far as I can see Leelenstein is mostly derivative of Leela. It uses a fraction of the games produced by Leela's self-play to get to a similar playing strength. Given that most of these games have also been used to train Leela, there is little reason to expect Leelenstein to be better. Basically Leela produces customized trainingsdata via self-play, ...


2

Python utilities for experimenting with Leela Chess Zero a neural network based chess engine: https://github.com/glinscott/leela-chess/ Here: https://github.com/so-much-meta/lczero_tools This allows you to run the network in Python on specific board positions via python-chess, and get policy/value outputs. (Works with pytorch, and is also able to run the ...


2

The weights for main networks are here. The best weights depends on whether you are using a cpu or gpu. For cpu, the best weights file is probably LD2 which is a small net trained off of data for t40. For strong gpus or longer searches, one of the recent t40 networks will be very strong.


2

After asking on the Discord as @Oscar Smith suggested in a comment, I got some results. The person I asked said that you could download good weight files for Leela from this webpage: http://lczero.org/networks/ This was also said: "nT40.T8.610 (used in TCEC Sufi): https://www.dropbox.com/s/rmcf0lf1hes10gi/256x20.T8-swa-610000?dl=0 and currently in use in ...


2

Part of this can be accomplished with Chessbase GUI. There is an option called cloud analysis. It is thougt to use engines in the cloud additional to your own engine. It ever could be used with two local engines, one determining the 1-n main variants, the other to evaluate the variants. A third engine could make proposals for answers to the main moves and so ...


1

Unfortunately, if the cheater is smart enough, he's never going to be caught! No matter how strong the cheat detector may be, if you only use the engine to help you out in one critical position, there is not enough data to get you caught. Finally, the only thing that can be proven is similarity between your play and the engine's play, which does not ...


Only top voted, non community-wiki answers of a minimum length are eligible