Leela Zero( A Neural Network engine similar to Alpha Zero)

Sort:
AIM-AceMove

seldepth: 22, nps: 42274, nodes: 38385, score: cuddn client on gtx 1060 20gb ram i7

Elroch

Could it be that as the network grows, a larger amount of CPU is required to play at an acceptable level, as well as providing the potential to play at a higher level?

drmrboss

I am 2100 bullet in lichess, leela is around the same rate as me. But I lose 3-0 in yesterday lol.

People who keep playing a lot of games vs engines usually get better results.

 

There is a bug in Leela where she see taking opponent pieces after discover check as illegal move

drmrboss

Read it twice or learn English.I said Lichess, here is my account.

https://lichess.org/@/drmrboss

drmrboss

Afaik, you are allowed to show games of other sites but you are not allowed to compare sites.

drmrboss

null

Leela reached another milestone. A big elo jump of Leela in these days. Leela ID 315 on 1060 GTX ,1+1 hit 3000 elo. 

Elroch

I was wondering when that was going to happen.

A great milestone. Chess engines are doomed!

[What was the time control?]

drmrboss
AIM-AceMove wrote:

Id 310. Finally i think blundering the queen is fixed in f5 trap line. Very impressive, feels like im playing bloody human.

And what is your elo of blitz over there? After winning Leela , did you reach 2500? Lol

I saw your blitz elo 2100 on yesterday., the same as me. wink.png. Btw , I lose leela 5-0 again, ID 320 play leela is  2700 in blitz, she clearly outplayed me. I cant beat 2700. 

Elroch
Elroch wrote:
ponz111 wrote:

can a chess engine over train?

I was referring to the key concept of over training in machine learning in general, whereby an AI fails to generalise well to new situations because it has learnt too much about irrelevant details of its experience. But your question is valid for Leela: I am not aware there is any chance it could suffer from this problem: it comes down to details of the training regime.

It is however, highly pathological for an AI to get worse performance from more data. More data is always better if it is relevant and used appropriately.

Confirmation that overtraining (also known as overfitting) was what caused the temporary hiatus in the unstoppable rise of LeelaZero to dominance of the chess world. wink.png

drmrboss
Elroch wrote:

I was wondering when that was going to happen.

A great milestone. Chess engines are doomed!

[What was the time control?]

Time control 1+1 mins

https://docs.google.com/spreadsheets/d/1XSJiCcQpCLv0fNwrUn7jXjdkZFU63YFEWpdXv6dSSg0/edit#gid=0

Elroch

Thanks. Do you mean 1 min + 1 sec per move?

I asked because I didn't think CCRL used that time control, but also because AlphaZero would have had a lower rating against Stockfish at such a time control than at its competition time control of 1 minute a move, even with its powerful hardware, according to the AlphaZero paper.

Note that the TPU hardware was much more powerful than a GPU as well, relevant to a fair comparison with Leela.

drmrboss
Elroch wrote:

Do you mean 1 min + 1 sec per move?

I asked because I didn't think CCRL used that time control, but also because AlphaZero had a lower rating at such a time control than at its competition time control of 1 minute a move, even with its powerful hardware, according to the AlphaZero paper. Note that the TPU hardware was much more powerful than a GPU as well, relevant to any comparison with Leela.

Because, there are about 4-5 new release of ID everyday. No one can test 40/40 or 40/4 conditions. It is still ok in testing at that time control. I think, stockfish was also testing 1 sec /game.(0.600 ms + some bonus I think)

Elroch

Of course: my main point was that AlphaZero is a bit nearer than it might appear!

drmrboss
This is extremely impressive play by Leela 320 ID, cuDNN on 1060GTX, 1+1 min time control, perfect chess 2017, stockfish 9 april version on i5 3.0 GHZ 4 cores
Quick  analysis showed 0 accurancies by both sides.
 
drmrboss
importantheart wrote:

Is she white or black?

White.

drmrboss
AIM-AceMove wrote:

I am capeable of beating title players (NM, CM, FM) on a good day.

What does this mean? 2000 rated player has chance to win 1 in 4 games vs 2200 player, 1 in 10 games vs 2400 player.

Winning a couple of games vs title player in bullet and blitz is nothing impressive, a lot of 2000 players can do that.

I won a few games vs GM . ( link in my wall).

AIM-AceMove

[COMMENT DELETED] changed my mind, no worth replying.

AIM-AceMove

COMMENT DELETED] made double comments

AIM-AceMove
[COMMENT DELETED]
AIM-AceMove
[COMMENT DELETED]