Skip to content

Latest commit



37 lines (22 loc) · 2.95 KB

File metadata and controls

37 lines (22 loc) · 2.95 KB

Leela Zero常见问题解答

Frequently Asked Questions about Leela Zero


Why doesn't the network get stronger every time


AZ also had this behavior, besides we're testing our approach right now. Please be patient.


Why only dozens of games are played when comparing two networks


We use SPRT to decide if a newly trained network is better. A better network is only chosen if SPRT finds it's 95% confident that the new network has a 55% (boils down to 35 elo) win rate over the previous best network.


Why the game generated during self-play contains quite a few bad moves

生成自对弈棋谱时,使用的MCTS模拟次数只有3200,还加入了噪声,这是为了增加随机性,之后的训练才有进步的空间。如果用图形界面(如sabiki)加载Leela Zero,并设置好参数与之对弈,你会发现它其实表现得并不赖。

The MCTS playouts of self-play games is only 3200, and with noise added (For randomness of each move thus training has something to learn from). If you load Leela Zero with Sabaki, you'll probably find it is actually not that weak.


Very short self-play games ends with White win?!


This is expected. Due to randomness of self-play games, once Black choose to pass at the beginning, there is a big chance for White to pass too (7.5 komi advantage for White). See issue #198 for defailed explanation.


Wrong score?

Leela Zero使用Tromp-Taylor规则(详见。虽然与中国规则一样贴7.5目,但为计算方便,并不去除死子。因此,结果与使用中国规则计算可能有所不同。不过,不去除死子并不影响模型的训练结果,因为双方会将死子自行提掉。

Leela Zero uses Tromp-Taylor rules (see Although its komi is 7.5 as in Chinese rule, for simplicity, Tromp-Taylor rules do not remove dead stones. Thus, the result may be different from that calcuated using Chinese rule. However, keeping dead stones does not affect training results because both players are expected to capture dead stones themselves.