supersharp77 wrote: ↑
Tue Jun 25, 2019 2:19 am
Robert Pope wrote: ↑
Mon Jun 24, 2019 9:03 pm
supersharp77 wrote: ↑
Mon Jun 24, 2019 6:40 pm
Laskos wrote: ↑
Mon Jun 24, 2019 7:53 am
What are these statements?
Lc0 T40 on an RTX 2070 GPU is stronger head to head than SF_dev on 8 i7 or Ryzen cores, from not too outrageous openings. What FIDE rating that means, you guess. I think it's impossible to _measure_ the FIDE rating of Lc0 T40 RTX 2070 in a pool of human FIDE players, as much too rarely a draw will occur. Aside that, Lc0 in a pool of regular engines, doesn't obey the Elo rating scheme (say CCRL Elo), as I posted here many months ago. So, besides the hardware and openings issues, the whole concept of an "Elo rating number" for Lc0 is theoretically superfluous.
Alpha Zero preprint and paper were a great scientific work which was confirmed by Lc0, and it doesn't need advocacy of some imbeciles.
Some Imbeciles? Like Who..You? Been testing LC0 for months and actually the engine rarely wins any games at all..It looks to be not a chess engine but only a NN list of played games with a basic search...My research shows a program that is erratic
disjointed and incapable of beating basic chess engines on CPU...It struggles with strategy, fumbles during the middlegame
and has little to no endgame knowledge..Program seems not to understand basic chess strategy. Performs like a random mover when out of its normal channels...And thats with me giving it the benefit of the doubt by adjudicating numerous games where LC0 should win because it is up material..but it struggles to finish games..It misses simple continuations to win games..3300 ELO 3500 ELO...4000 ELO...? Not in the games I'm viewing...A semi random mover with a HUGE NN opening book...that needs huge speeds to be effective at all...Lost to Bikjump 1.8 x 64 yesterday and King 3.32
I don't think anyone disputes that Lc0 plays much worse on a CPU than a GPU. That's hardly a revelation. And has nothing to do with AlphaZero, which never played on just a CPU, that I am aware of.
So back to my original premise..was the Alpha Zero Claim of 4000+ ELO with 24hr of Self Play Games a Proper Claim?
and What is the Actual Strength Of Alpha Zero and Or LC0? Thx AR
The Alpha Zero claim of the 24hr NN of 4000+ elo is not knowable by us, unless there is some data of games I have not seen.
And the actual strength of Lc0 and Alpha Zero depends on what NN, and what hardware is being used.
I believe the Alpha Zero that played Stockfish had 9 hours of training. And depending on what match, and games you accept to rate. AZ was between 44 and 78 elo better then Stockfish 8 by performance. Using this hardware Stockfish 8 Hardware 44 CPU cores, Syzygy endgame tablebases, and a 32GB hash size vs Alpha Zero's four TPUs.
If we assume the numbers here are more or less accurate for this speculation.
Stockfish 8 64-bit 3302 CCRL 40/40 rating. Equivalent to 40 moves in 40 minutes on Athlon 64 X2 4600+ (2.4 GHz)
Stockfish 10 64-bit 3389 CCRL 40/40 rating. Equivalent to 40 moves in 40 minutes on Athlon 64 X2 4600+ (2.4 GHz)
What is the playing strength of Stockfish 8 on a 44 core system vs a 1 core Athlon 64 X2 4600+ (2.4 GHz)
Lets assume a win ratio of 95% or 1 draw every 10 games. That would give a rating for Stockfish 8 on the 44 core system of 3814 elo. And at best AZ could be 78 elo better then Stockfish 8 on the played hardware. So AZ's 9 hour NN could be as strong as 3892 Elo.
Now could AZ's 24 hour trained NN be 4000+ elo. Sure, it is possible and probable.
Stockfish 10 by CCRL is 87 elo stronger then Stockfish 8. And it is also possible that Stockfish 10 is stronger then AZ. And it is also probable that Lc0 is stronger now then AZ. Since Lc0 is stronger by performance then Stockfish 10.
But speculation is not real games, but this is the best data we have.