Leela Zero (Stockfish without human knowledge), like Alpha 0

Sort:
drmrboss

 

Leela is back with massive elo gain in the last two days (300 elo in self play which is probably about 200 elo). 

null

drmrboss

Leela already passed 2000 +!!

 

Note that all tests have fairly big error margin,
also LCZero scales better than engines usually with more time
!1min+1s, 1knps LC0 id80 against various engines, 1822 Elo (+68 over id69) (CCLS)
!48s+0.2s, 5knps LC0 id79 against various Stockfishes, 2139 Elo (+139 over id67) (zz4032)

drmrboss
nighteyes1234 wrote:

I thought it was the worst its played in awhile. Same old Berlin 1500 rated opening. Missed a mate in 2. Said it had a 45% chance of winning, and post-analysis eval showed me up 5.5, thanks largely in part to sacrificing a bishop on my h3 pawn.Went from 42% chance(when I was up 8.5 and then easy M2) to 1% in one move.

 

If I play Bc4 instead of Bb5 Ruy , it goes for same ole Philidor defense and its '53%' chance of winning as black. 2000? no way....Im thinking more like 1500. Actually I dont know any 1500 who plays Philidor and thinks black is way ahead.

In play.lczero.org website, Ieela's performance will be significantly bad if there is multiple requests from multiple accounts . You will see real leela strenght if you have a good NVIDIA card, with blitz or rapid setting!

nighteyes1234
drmrboss wrote:

In play.lczero.org website, Ieela's performance will be significantly bad if there is multiple requests from multiple accounts . You will see real leela strenght if you have a good NVIDIA card, with blitz or rapid setting!

 

I'll be patient . The id#82 didnt completely mess up the berlin  this time and tried to play defense. Do you think it thinks Berlin is the best defense against e4 or do you think it will play the sicilian? Didnt A0 evolve over several openings?

drmrboss

It will take a long road for her to see the best opening. Certain amount of games are needed for leela to learn.  For example, after playing 10,000 games with berlin, she saw 40% score with black., Then she will find something else. If she found Ruylopez scoring 50% for black for she will choose it. By this way she will choose the best opening. In 10 games of SF vs A0, A0 chose the same opening. 

AstroHorse7

Leela is now 3219

Elroch

Very nice progress. How far can she get?

AstroHorse7
Elroch wrote:

Very nice progress. How far can she get?

In a year she will be better than Stockfish.

Elroch

Now that's a nice goal!

drmrboss
AstroHorse7 wrote:
Elroch wrote:

Very nice progress. How far can she get?

In a year she will be better than Stockfish.

Well, current elo of leela on 1060GTX is like 3200 while we assume 4 cpu on stockfish is like 3500. So there is about 300 elo gap.

The problem is that the top elo are extremely hard to get. (for comparison- stockfish needed 25000 patches and 5 years of 880 million games to squeeeze the top 300 elo from 3200 to 3500).

 

And also Alpha Zero also stalled at 3300 elo since after 14 million games to 44 million games, so the chance of passing 3500 elo on common hardware like 1060 GTX is still possible but  small.

drmrboss

null

This is today results from Lczero, 3200+ rating on 1070GTX Ti. 

drmrboss

null

 

Leela Zero 20x256 net(exact replica of A0)  vs  Alpha Zero.

 

Initially Leela group's training method seemed better, but at current level (after 19 million games), Leela training method seemed about 150 elo weaker than A0.

 

 

Elroch

That is not a concern. The slope looks great (slightly better than A0 above 3000 for some reason, just a bit later).

AstroHorse7

Leela is now 3400-3500!

number-0

yeah if alpha zero had a kid it would be leela.

so leela is kid/infant of alpha zero

drmrboss

She is doing well in CCCC, as of today, she is No 2. now. null

JamesAgadir
DamonevicSmithlov a écrit :

I believe Alpha Zero scored 64% vs Stockfish. Was that Stockfish version about 3500 or so? If that's the case Alpha Zero is about (or was at that time) 3600 - 3650ish perhaps? It seems that if chess.com does another computer tournament at the end of the year then Leela (Leela "learns" the same way as Alpha Zero, right?) should win it at the rate it's improving.

The problem with that is that they gave Stockfish a major disadvantage on processing power. Stockfish was probably better then alpha zero, but if you give one racer a small clio and the other one a ferrarri you won't find out the better race driver

nighteyes1234

Leela is slowly getting more interesting. My engine(SF framework) didnt see any sizable benefit  to this attack vs the French for 20 moves before it saw the correct eval right away. Most of the time my engine was in a blackhole.

 

drmrboss

This blog post is very good in setting up Lco properly.

http://blog.lczero.org/2018/09/guide-setting-up-leela-on-chess-gui.html?m=1