George Tsavdaris wrote: ↑Mon Sep 17, 2018 10:32 pm
For an interested Chess player, that has already asked in vain around 15 times in the CCCC chat, where is the PGN of all the stage 1 games?
Thanks for diggin' out that link. If someone is interested I will clean it up (e.g. 17 pgn headers are not necessary), add round tag numbers
and remove the pv comment for readability.
All cleaned up and ready, but there is at least one game missing:
Rd 46. Laser-LC0 is not there!
Also Rd. 28 Crafty - Laser has only two unterminated games, while I thought the rules
require a replay until there is a result.
There was already a predecence for this, as Ivanhoe - LC0 was started three times
until there was a result.
So we have 569 games (including 19 unterminated games) = 550 (should be 552 terminated ones, 46*12)
jkiliani wrote: ↑Mon Sep 17, 2018 4:33 pm
Lc0 now has a training run using tablebase rescoring for training, rather soon we should be able to see improvements in endgame play from this.
And you still keep "0" in the name, what a joke...
Yes. They do that, to annoy you. And it's good.
But that's a side experiment, right?? The main nets better continue to be "0".
I see it's test30 that has TB cheating. If that goes over to the main net, it'll be a ridiculous decision totally against the written first goal of the Lc0 project.
jkiliani wrote: ↑Mon Sep 17, 2018 4:33 pm
Lc0 now has a training run using tablebase rescoring for training, rather soon we should be able to see improvements in endgame play from this.
And you still keep "0" in the name, what a joke...
Yes. They do that, to annoy you. And it's good.
But that's a side experiment, right?? The main nets better continue to be "0".
I see it's test30 that has TB cheating. If that goes over to the main net, it'll be a ridiculous decision totally against the written first goal of the Lc0 project.
I first goal of Lco to follow A0 was almost complete since 10xxxx series.( Simulation tests showed that Leela was very close to A0, within 50 elo level, A0 and SF 8 were history of 1 year ago anyway)
Second goal is to fight for No 1 and now she is among big 3 level in CCCC.
Series of 20xxx , and 30xxxx networks are aiming to be No 1. (Leela should be no 1 in 2019)
jkiliani wrote: ↑Mon Sep 17, 2018 4:33 pm
Lc0 now has a training run using tablebase rescoring for training, rather soon we should be able to see improvements in endgame play from this.
And you still keep "0" in the name, what a joke...
Yes. They do that, to annoy you. And it's good.
But that's a side experiment, right?? The main nets better continue to be "0".
I see it's test30 that has TB cheating. If that goes over to the main net, it'll be a ridiculous decision totally against the written first goal of the Lc0 project.
I first goal of Lco to follow A0 was almost complete since 10xxxx series.( Simulation tests showed that Leela was very close to A0, within 50 elo level, A0 and SF 8 were history of 1 year ago anyway)
Second goal is to fight for No 1 and now she is among big 3 level in CCCC.
Series of 20xxx , and 30xxxx networks are aiming to be No 1. (Leela should be no 1 in 2019)
I disagree.
A0 was known to scale well and to be weak at blitz.
I read that it is the opposite with LC0 so LC0 has clearly something different from A0.
Edit:
I also disagree that Leela should be number 1 in 2019.
Other programs also improve and in any case I cannot define Leela to be number 1 based on unequal hardware competition.
Edit 2:I also disagree that now leela is among the big 3 level in CCCC.
I see
1)Stockfish 11/15
2)Houdini 10/15
3)Komodo 9/15
4)Lc0 9.5/16
I guess stockfish houdini and komodo are the big 3 and so far lc0 is lucky.
Jhoravi wrote: ↑Wed Sep 19, 2018 4:19 pm
Disallowing update between stages will just show a similar result from stage 1.
Not exactly similar. She will definately get better due to
1. Ponder off (Pondering On hit her Time Management algorithm and her TM was terrible with in stage 1. I guess she is one that suffer most with Ponder ON)
2. She is generally drawish. ( her scores will relatively improve more cos other engines will also draw more in stage 2).
So far, she is the one that can positionally outplay Stockfish and almost got the chance to win vs SF.
[pgn][Event "CCCC 1: Rapid Rumble (15|5) Stage 2"]
[Site "Chess.com"]
[Date "2018.09.18"]
[Round "?"]
[White "Lc0 17.11089"]
[Black "Stockfish 220818"]
[Result "1/2-1/2"]
[BlackElo "3283"]
[ECO "B22"]
[Opening "Sicilian"]
[Time "22:30:35"]
[Variation "Alapin, 2...e5"]
[WhiteElo "3300"]
[TimeControl "900+5"]
[Termination "normal"]
[PlyCount "362"]
[WhiteType "human"]
[BlackType "human"]
Thanks to everyone interested in the CCCC. If you are interested in getting involved, let us know. We are looking for an official TD/manager, a CuteChess developer, and chat moderators.
Please fill out this form and we will get back to you. Cheers!
Jhoravi wrote: ↑Wed Sep 19, 2018 4:19 pm
Disallowing update between stages will just show a similar result from stage 1.
I understood that updates are not allowed and it is the same lc0 as stage 1 and also the same stockfish as stage 1.
Did I understand wrong?
In any case I think regardless of time management that ponder off means effectively shorter time control and shorter time control is going to help lc0.
Inspite of it I guess lc0 is going to get only 4th place.
it now shares place 3-4 with komodo with 10 out of 17.
