Some Elf games are/were (I think they are being thinned/phased out recently) used in the LZ training. I'm not sure if these are real Elf engine+weights self-play games released in some huge dataset by Facebook, the LZ engine with converted Elf weights playing against itself on LZ volunteer machines, or LZ engine with Elf weights vs LZ engine with LZ weights playing on volunteer machines, but suspect it's the middle one. Can anyone confirm? When these Elf games were added to the training pipeline LZ improved very fast initially as it absorbed the Elvish wisdom. You can see latest LZ has a more Elvish style; not so obsessed with 3-3 invasions of 4-4 but approaches them too, the 4 corners then 4 approaches fuseki liked by Elf is now also played by LZ (almost exclusively in the vs elf test matches I looked at, gets kinda boring). Also Elf is somewhat more human in not hating side extensions so much or playing hoshi pincers (kinda retro Japan style, you won't see Koreans do that since the 90s!). Whether LZ would have evolved like this without the influence of the Elf games we can't say, but I think it's fairly safe to say it has speed up that process.

The nets are not trained on the games themselves but on search results in individual positions. So training on Elf vs LZ games would not be much different to training on some Elf games and some other LZ games (except slightly more variety in the types of positions occurring, and in game results for the value net). What matters is which net was used in the search at each position (whose turn it was).

The Elf games in the training are LZ selfplay with Elf weights AFAIK (the other options are not readily available).

The Elf vs LZ idea sounds most interesting for value net training but also very risky, may introduce some biases: game results would not reliably reflect the position itself anymore but also the strength difference.

In another thread, there was a very heated discussion about LZ 0.14 being stronger or weaker than LZ 0.15

But now, there's LZ 0.15.9999 (not a gag, you can download it HERE)
Discussion about V0.15.999 and V0.15.9999 HERE

V0.15.9999 is a test for V 0.16, and... it seems really stronger than V0.15

20 game match between
LZ 0.15.9999 network #184 and LZ 0.15 network #184
(3 min per side and per game, no pondering, GPU : 1x1080)
LZ 0.15.9999 wins 14-6
Big win...

Edit : another 40 game match 0.15.9999 v 0.15 (#184, same conditions as before)
LZ 0.15.9999 wins 22-18

So, after 60 games : LZ 0.15.9999 wins 36-24
Not bad


If someone wants the games...

It's not suprising it is stronger : it's faster.

Look at the difference in playouts you got in your matchs

"visit" parity test 0.15.9999 vs 0.15: 193 wins, 191 losses

Nothing interesting ...
P.S. If someone wants the games...

The official V0.16 just came out !
V0.15.9999 or V.0.16 is faster than V0.15, so, it's stronger than V0.15 at TIME PARITY, not at visits parity (your test is at visits parity...)

If you run a match at TIME PARITY, I bet you'll see a big difference !

How much faster is the question. I just did a little test on my GTX 1060 and LZ #157 in 10 seconds went from 4k playouts to 5.4k. So that's about 35% speed increase, nice!

Someone posted this at github about the improvement LZ 0.15 --> LZ 0.15.999



So, it seems you're about right !

And time parity is the correct way to measure.

However ENOUGH time (realistic time) can be critical also. A program that might be better than another at unrealistic short amounts of time but dos not improve as much when given a more realistic amount of time as the second program is not necessarily the stronger at REALISTIC time.

We really should be making some decisions about this. What would be realistic for human blitz games? What would be realistic for fast (but not blitz) human games? What would be realistic for normal speed human games?

New network (185)
20 game match between #157 (best 192x15) and #185 (best 256x40)

LZ 0.16, 5 min per side and game, GPU 1x1080, twogtp 1.4.10

#185 wins 12-8 (6 wins as B, 6 wins as W)

Average length 227 moves , almost exactly 2 sec/move for both sides.

I'll upload the games if someone wants them.

"True" time parity test - v16 vs v15 (3s per move, 2x1080ti, #184):

71 : 29

I'm following Leela's progression with delight. Does anyone know what its current rating against human pros on "normal" hardware (let's say one GTX 1080) would be? Is it better than the #1 human on those conditions? If so, by how much?

_________________
Sum ergo non ero.

AND normal time. A human player could do little to no reading at 2 seconds/move.

Considering that even a few months ago, LeelaZero was beating human professionals with 2-3 handicap stones, which it was not even built for, on consumer hardware. (Since then, there has been various improvements to handicap play)

I would think that LeelaZero is well above any human on a GTX 1080. It is possible that you can find a blind spot with a ladder or similar though, depending on the time settings.

I run two games with equal thinking time (20 seconds per move) between Phoenixgo(W) and Leela zero(B) with weight 186. LZ lost both games.

LZ weight 185 is better against Phoenixgo.

"True" time parity test - #186(black) vs #157a (3s per move, 2x1080ti).

Empire strikes back, part1:

19:31

P.S. Games #22 & #34 are the top of artificial intelligence .
What/where are we doing/going?

to be continued...

What is your GPU ?

GTX1050 with 2G memory. In the middle game, 20 s means about 1500 playouts for LZ and more playouts in the end game.

"True" time parity test - #186 (white) vs #157a (3s per move, 2x1080ti).

New hope, part2:

29:21

Finally: #186 vs #157a : +48-52=0

P.S. I can not explain the results of my test. What is wrong with black colour? V16, "--precision half" option, too little games...?