Leela0 (LC0) cpu Initial Observations

[supersharp77]

Recieved my w64 Laptop about a week or so ago and after updating the software and transferring data and my chess engine files from my w32 laptop I was in the mood to pick up where I left off on my ongoing w32 tests and get the lowdown on some of the more recent developments concerning w64 and the ongoing debate on Neural Networks...after loading Shredder gui Fritz gui and my beloved Lokasoft (Chesspartner) gui I was in the mood for some w64 testing.......after loading some engines I had waited years to load I was ready..locked and loaded....and last but not least..Of course I had to try and get LC0 (Leela0) A shot! (Had tried in w32 with no success previously) Spent a day or two reading and found some leads on what to do then downloaded some new cpu files and tracked down some old files downloaded last year...found the secret of clicking on the "client" file first with the unzipped Neural Network in the file and YES!! after registering on the network the LC0.exe file would then work. Set up 5-6 different LC0 engines and placed them in the various gui's set up some tourneys and started testing Lot of time losses using my standard time controls of 5 minutes +5 seconds....also 2 minutes = 12 seconds not long enough 3 minutes +20 seconds or 5 minutes +20 seconds might be just long enough for LC0 Leela0 based on my game observations Engine seems more suited to longer time controls...Early losses to Bikjump,Embla (time) Vapor (3 on time and by the position) Houdini (Under 30 moves) Robbolito & Ivanhoe...Engine thinks no more than 4-5 plies with an occasional 7-8 plies search observed. Based on my observations engine is poor on tactics ('I think it missed some combinations') and not too good in the endgame either...Opening play is spotty with occasional flashes and the unusual and sometimes Genius Level Positional Play..Engine castles queenside more than most and sometimes castles manually. Given enough time and the right set of positional circumstances Leela0 (LC0) is capable of exceptional positional play...Seems very adept at sacrificing pawns for positional advantage (1-2 pawns). Have not observed any piece sacs as of yet...My opinion and based on my observations of the playing style is that if millions or (billions) of opening positions are available using the network and if the speeds (speed keeps coming up in reference to the engine and networks) Are as fast as I think they are...they would present an ""unlimited opening database"" and depending on how deep they go (games database..aka neural network) are you playing against the chess engine or the neural network? This presents a similar problem presented by playing against Brainfish with a multmillion games cerebellum book or something like that...Based on what I have observed so far no way this engine is anywhere close to 2800 at fast time controls (2200-2400 tops) Of course as the time goes up....engine strength would probably increase exponentially. Meanwhile I will keep tweaking and observing Leela0 (LC0) in my ongoing tournaments and will publish some result and games when more data becomes available

[dkappe]

Can I ask which network files you are using? If you are using a full size network on cpu, it will do quite poorly. Give some of my distilled networks a try. The 112x9 network is especially good on cpu. I’d estimate it as about 2800-2900 ccrl.

https://github.com/dkappe/leela-chess-w ... d-Networks

[supersharp77]

Can I ask which network files you are using? If you are using a full size network on cpu, it will do quite poorly. Give some of my distilled networks a try. The 112x9 network is especially good on cpu. I’d estimate it as about 2800-2900 ccrl.

https://github.com/dkappe/leela-chess-w ... d-Networks

Downloaded from the LC0 github site based on the authors recommendations zip file was 45 mb..opened to 56 mb another one was 256 unzipped....In my tourneys engines have gui books ar they can use the engine preloaded books so a chess engine without a opening book could be "flying blind"...

ps without a book Leela0 falls behind and quickly gets into severe time trouble...especially at blitz time controls 5 min +5 sec
or 2 min +12 seconds...

[dkappe]

On my gpu (a 1070) the 256x20 nets do about 4000 nodes per second. On a cpu, we’re talking 30-40 nodes per second. The smaller 112x9-se net will do many 1000’s of nodes per second on cpu. It plays about 300 elo stronger than a full sized net on cpu. (Remember, the net weights are the pb.gz files.)

It must be extremely frustrating for you trying to use leela full sized networks on just cpu. Do yourself a favor: try one of the smaller and faster networks from my page.

[M ANSARI]

Using Lc0 on a CPU is like running SF on an old 386 chip, it just won't do well. You need to run it on a good GPU for it to become very impressive (and it is impressive). There are some dramatic tactical weaknesses that come out every once in a while and to be honest, I have no clue why as it can also play some incredibly deep tactical play. My guess is there are still some tweaks in the software and how the engine works to patch up some of the holes. On an RTX 2080 Ti, LC0 repeatedly scored about 90 ELO stronger than the latest SF Dev running on 6 cores ... and that is including the losses caused by unpatched weaknesses. Of course this is without books as with books it seems to blunt Lc0 advantage.

[supersharp77]

Using Lc0 on a CPU is like running SF on an old 386 chip, it just won't do well. You need to run it on a good GPU for it to become very impressive (and it is impressive). There are some dramatic tactical weaknesses that come out every once in a while and to be honest, I have no clue why as it can also play some incredibly deep tactical play. My guess is there are still some tweaks in the software and how the engine works to patch up some of the holes. On an RTX 2080 Ti, LC0 repeatedly scored about 90 ELO stronger than the latest SF Dev running on 6 cores ... and that is including the losses caused by unpatched weaknesses. Of course this is without books as with books it seems to blunt Lc0 advantage.

"Faster is better" Sounds great in theory.....but have been hearing these same theories for years..for example just watched ASM fish with w64 struggle to draw with Octochess w64.. trust me faster is not always better my friend...just looked at some of the coding inside LC0 (not much inside) great code creates great chess engines....like I proposed before how are we sure we are playing against chess engine and not the games database? And when does the multimillion game database end and the chess engine begin? Like mentioned above engine is great at certain positional concepts but misses some obvious tactical motifs..These engine tests are ongoing...."Put up or shut up time draws near".....

[dkappe]

"Put up or shut up time draws near".....

Not sure what that means. Nothing draws near. Every month or so new features are added and newer, stronger nets come along.

If folks are running misconfigured lc0 or lc0 on cpu, I’m more than happy to help.

[supersharp77]

"Put up or shut up time draws near".....

Not sure what that means. Nothing draws near. Every month or so new features are added and newer, stronger nets come along.

If folks are running misconfigured lc0 or lc0 on cpu, I’m more than happy to help.

engine seems to be running and playing ok.....time losses at various time controls are becoming annoying...any settings
ideas to stop time losses at 4min + 20 sec or 5 minutes +20 sec or 5min +5 sec or 2 min + 12 sec /move? Thx AR

[Redshift]

Have you tried what has already been suggested, switching to a smaller net size?

[Hai]

"Put up or shut up time draws near".....

Not sure what that means. Nothing draws near. Every month or so new features are added and newer, stronger nets come along.

If folks are running misconfigured lc0 or lc0 on cpu, I’m more than happy to help.

engine seems to be running and playing ok.....time losses at various time controls are becoming annoying...any settings
ideas to stop time losses at 4min + 20 sec or 5 minutes +20 sec or 5min +5 sec or 2 min + 12 sec /move? Thx AR

=GPU is to weak