The graph is misleading because of a bug that has now been solved (though, the graph still shows old data). There is around 300 "self-play elo" difference between network id 55 and the previous one, which the graphic does not show. The progress hasn't stalled at all.
# PLAYER : RATING PLAYED W D L (%) D(%) OppAvg OppN OppDiv
184 Leela Chess Zero Gen 33 (ID 55) x64 : 1632.2 139 92 9 38 69 6 1441.8 25 24.9
200 Leela Chess Zero Gen 30 (ID 44) x64 : 1506.3 200 106 22 72 59 11 1427.0 25 25.0
208 Leela Chess Zero Gen 24 x64 : 1401.3 200 84 22 94 48 11 1427.0 25 25.0
213 Leela Chess Zero Gen 20 x64 : 1317.8 200 64 27 109 39 14 1427.0 25 25.0
226 Leela Chess Zero Gen 16 x64 : 1229.3 200 47 26 127 30 13 1427.0 25 25.0
228 Leela Chess Zero Gen 14 x64 : 1185.3 200 42 20 138 26 10 1427.0 25 25.0
232 Leela Chess Zero Gen 12 x64 : 1094.2 250 64 26 160 31 10 1278.6 49 40.8
244 Leela Chess Zero Gen 10 x64 : 860.1 92 53 11 28 64 12 654.5 23 23.0
247 Leela Chess Zero Gen 8 x64 : 791.6 92 45 17 30 58 18 654.5 23 23.0
252 Leela Chess Zero Gen 6 x64 : 597.4 92 31 18 43 43 20 654.5 23 23.0
256 Leela Chess Zero Gen 4 x64 : 369.0 150 43 18 89 35 12 623.1 15 15.0
Yes but after the bug is fixed i see that the networks after 55(excluding ID58 that shows 0 games) have a decline in the rate of the increment of the ELO, as also even a decline in ELO the latest ID57!
David Xu wrote:LCZero possesses a "-noise" command that applies Dirichlet noise to its move selection, thereby causing randomness in its play.
Quick correction: --noise does not cause randomness in its play, it just introduce some randomness in the search, but the moves introduced are evaluated and discarded if they are found to be not good enough.
The randomness in its play that you see in self-play training games is caused by another thing, temperature=1
Just a question. Should the noise option be used at all, or will it hurt a bit when using it in match play vs. other programs with given start positions repeated from both sides?
David Xu wrote:LCZero possesses a "-noise" command that applies Dirichlet noise to its move selection, thereby causing randomness in its play.
Quick correction: --noise does not cause randomness in its play, it just introduce some randomness in the search, but the moves introduced are evaluated and discarded if they are found to be not good enough.
The randomness in its play that you see in self-play training games is caused by another thing, temperature=1
Just a question. Should the noise option be used at all, or will it hurt a bit when using it in match play vs. other programs with given start positions repeated from both sides?
Noise DOES cause randomness in play, just a lot less than temperature. If the scenario you're using doesn't need any randomness in play, you don't need to use --noise either.
27.Qc4????
It definitely has some serious bugs! Can you post it in the lc0 forum in order to inform its programmers about it? Because in the games i played it also dropped a Bishop for nothing too. It seems general and serious.
After his son's birth they've asked him:
"Is it a boy or girl?"
YES! He replied.....
George Tsavdaris wrote:
Yes but after the bug is fixed i see that the networks after 55(excluding ID58 that shows 0 games) have a decline in the rate of the increment of the ELO, as also even a decline in ELO the latest ID57!
Yes I'd love to hear why ID 60 is so much weaker than ID 55.
Is it another bug? Do the bug fixes lower elo before it climbs again? There seems to be a very regular "one big step forward" and then "lots of little steps backwards".
But this is a training game not a real game. Those always have been terrible and will be for quite a while since temperature is used. (which introduces A LOT of randomness so that one gets different games even with such a huge number of games, also of course to possibly escape local maxima)
Why people keep posting training games is beyond incredible... it's been said a million times already that they are NOT representative of real strenght.
ID 55: 61.0/200
ID 69: 72.5/200
ID 80: 89.0/200
Yet, a (somewhat serious) bug was discovered yesterday and major changes are under way, which when combined with the engine improvements that are about to be released, will make a huge jump in strenght in the days to come... we'll see how things develop.
