leela chess zero is a hype

with that play...i can guarantee you leela has to play itself for a century to hit a 3000... I think that self learning thing Is another hype

notOldenough i can't beat leela u say,but bet mt stockfish 7 can... but seriously speaking i think i stand a big chance

Thanks, EscherehcsE.  My network is 20758.  Its "training ELO" is 4234.54.  I think it was the highest  ELO when I downloaded it, but some more recent networks have higher ELOs.  Is that ELO column the best way to judge the relative strengths of these networks, and if not, is there a network you would recommend for entry-level GPUs (such as my 1050 TI)?

Best  Leela network ID is 11248, engine ID is 18.1.

20xxx series and 30xxx series are under training. They are about 300 elo weaker than 11xx series.

Yep, I agree with drmrboss about the 11248 network. I just went to this page:

https://github.com/LeelaChessZero/lc0/wiki/Training-runs

That page mentioned three of the best networks, so I just self-played those three to get the best one. Networks 11248 and 11250 are close to identical. Maybe 11248 is the best, but 11250 is verrry close. Network 595 is much weaker than the other two.

To me, the "Training ELO" numbers seem to be highly inaccurate and just about meaningless. I pay no attention to those numbers.

I just downloaded and tested the 11176 network with a repeat of the 10-game bullet match with Hermann.  Lc0 won this time, +5 -3 =2.  I'll try 11248 next.

Just to mention the basic advice on statistics - 10 games isn't nearly enough to get an accurate feel for ratings. You'd have to make at least several hundred games to get any accuracy at all. (And preferably even more games than that...)

Leela-11248 beat Hermann +4-2=4, but there were some very strange games.  Leela seemed to get the advantage easily but would sometimes throw it away for no reason.

There have been a number of games like this since I started playing with Lc0 last night.  Maybe it's because of the 1-minute time control.

I can think of two possible reasons for this strange behavior:

1) Perhaps Leela needs some tablebases. Try adding a 5-piece Syzygy tablebase set. You'll have to add the Syzygy option as a command-line option.

2) Perhaps Leela is running out of time near the end of the games. (In my early tests, Leela was losing a significant number of games due to time forfeits.) Try making these two changes to the Leela engine configuration:

Scale thinking time: 0.500000
Move time overhead in milliseconds: 10000

No doubt.  I'm just playing around with it, really.  I suspect, just from watching the games, that the difference between Lc0 and Hermann (on my machine, at least) will widen with longer time controls.

Yikes!  Are you sure there isn't a typo in there?  Leela is losing every game, badly, with those parameters...and it's over.  Leela lost 0-10.  I'm going to try 1000 instead of 10000.

No typo. One-half for the scale thinking time, and the maximum value of ten thousand for the move time overhead. At least these numbers worked for me. Maybe these numbers work differently depending on whether a GPU card is used or not? idk

I'll have to play around with it some more.  Dropping the move time overhead to 1000 improved Leela's score slightly, but she still lost +1 =1 -8.  I'm starting to think the parameters might have to be hand-tuned for a particular hardware/network configuration.  I should probably try tweaking one parameter at a time.

Anyway, I'm having fun.  Thanks for your help, EscherehcsE.

I was wondering what was wrong.  I have a Ryzen 5950 128 GB Ram and an NVidia 3080 graphics card.  I downloaded the CUDA version but it was not performant.  Then I downloaded a 30 block network and set the engine parameters correctly and now it's a beast.  It processes 200K positions per second which for Leela is outstanding.  Of course, Stockfish analyzes 35MN/s.  Stockfish 14 is my go-to engine, but Leela is incredibly strong and useful.