Acronyms in Leela Zero

Finishing watching the anime “Hikaru no Go” let me wanting two things so badly: a traditional Japanese floor Go board and a world no. 1 ghost player that can play and teach me every night. The first craving is easy to satisfy so long as I part with some money, while the second one seems ridiculous and impossible. Until…

I learned that Chinese national Go team uses an AI designed by Tencent called “Jueyi” - it’s literally the head coach there. Unfortunately it is not available for public use. Then my research lead me to Leela Zero, and… to me it is just the “Fujiwarano Sai” that I wanted! I have a relatively good PC for professional use - Xeon E5-2670 64GB GTX 770. Even limited at 2 seconds per move thinking time and 100 playouts, it is stronger enough to teach me a whole lot. I am amazed by how clean and concise the software is - I use Sabaki and Leela Zero combination, downloaded from their offical web page. And I am impressed by it’s strength, especially the fact that the software was published less than a year ago.

It seems Alpha Go Zero’s approach - not knowing anything except rules - is the right way to build up a world no. 1 player.

I have some questions though, that I hope people can help answer. There are so many acronyms in the command lines. I can figure out some by diligently researching on the Internet, but there are those that I just cannot figure out. Particularly in this screenshot:

There are so many percentages. What is “W U V N %”? Also there is this “MC winrate”. They are all different. I assume in this older version, “visits” means “playouts” in later version. (correction: I just realized visits and playouts are two different things. I need to do more research on that…)

I cannot find an official documentation of Leela Zero. If anyone has a link to a detailed documentation, could you please post?

Many many thanks! …and I’ll go back to my training with LZ - another acronym! (that I figured out myself)

Abbreviations are not the same as acronyms. Letters used as parameters are usually called flags or commandline options.

This screenshot is leelasabaki running with the old leela, not leelaz. There is some documentation on the leela site about what the stats mean. AFAIK, MC winrate is winrate for the current position from Monte-Carlo playouts, as opposed to the estimation from the value network. I don’t think Zero does any MC playouts and there is only one network, so stats for LZ should be simpler.

Thanks.

Here’s a sample of a typical move that I just got from testing LZ in Sabaki:

"

  1. sabaki-genmovelog

  2. = #sabaki{…}

  3. D4 -> 14 (V: 53.80%) (N: 69.30%) PV: D4 D16 Q4 R6 R14

  4. Q4 -> 5 (V: 54.03%) (N: 12.91%) PV: Q4 D4 D16

  5. D16 -> 5 (V: 53.90%) (N: 13.18%) PV: D16 D4 Q4

  6. 3.7 average depth, 6 max depth

  7. 16 non leaf nodes, 1.50 average children

  8. 25 visits, 8957 nodes, 24 playouts, 23 n/s"

It has V rates and N rates. I guess something like “win rate” “neural rate”.

Also playouts and visits are clearly two different numbers.

Thanks for your response.

Just a guess, but I think what you’ve got there are (V) the win rate (53.8%), which is the estimated win rate at the end of the principal variation, (N) the network evaluation of the current position, and (PV) the principal variation.

Maybe check out Lizzie, for a better interface, and Leela Zero?

But of course, remember…