Page 9 of 27

Re: LCZero update

Posted: Fri Mar 16, 2018 8:16 pm
by jkiliani
noobpwnftw wrote:One thing to mention about the NN approach is when it works it works, but when it doesn't, you usually can't fix it. There is an example in recent Go engines recreated from AG paper that would fall for some particular position every single time and the amount of training samples wouldn't be enough correct it, not to mention that doing human intervene contradicts the philosophy of "learning on its own", so maybe this is what they choose not to tell you and we probably will never see any A-series gets published for public use.

Sounds funny when you think of their medical plan, do the AI need to kill enough patients before learning not to, if so, how many, a zillion?
Care to share what that specific Go engine and the specific weakness is?

Re: LCZero update

Posted: Fri Mar 16, 2018 8:31 pm
by noobpwnftw
jkiliani wrote: Care to share what that specific Go engine and the specific weakness is?
Fine Art, discussions in Chinese here.

Re: LCZero update

Posted: Fri Mar 16, 2018 8:39 pm
by jpqy
Is someone using Lczero already into FritzGui ,cutechess or any Gui already?

Re: LCZero update

Posted: Fri Mar 16, 2018 8:45 pm
by jkiliani
noobpwnftw wrote:
jkiliani wrote: Care to share what that specific Go engine and the specific weakness is?
Fine Art, discussions in Chinese here.
OK thanks, I guess I can't follow that discussion then. Thought you were talking Leela Zero + ladders

Re: LCZero update

Posted: Fri Mar 16, 2018 8:55 pm
by noobpwnftw
jkiliani wrote: OK thanks, I guess I can't follow that discussion then.
The discussion was about the lack of variations played by Fine Art might because of some over-fitting and the engine will always fall for it as shown by the images, several games were identical and was done intentionally just to prove it.

Re: LCZero update

Posted: Fri Mar 16, 2018 9:00 pm
by pferd
jpqy wrote:Is someone using Lczero already into FritzGui ,cutechess or any Gui already?
I am running it in xboard via:

xboard -fcp "./lczero --weights get_network -p50000 --noponder" -fUCI and can play games against or let it play against other engines.

Does anybody know how strength increases when the number of playouts is raised?

Re: LCZero update

Posted: Fri Mar 16, 2018 9:05 pm
by CMCanavessi
jpqy wrote:Is someone using Lczero already into FritzGui ,cutechess or any Gui already?
I'm using it in Arena.

Re: LCZero update

Posted: Fri Mar 16, 2018 9:08 pm
by CMCanavessi
pferd wrote:
jpqy wrote:Is someone using Lczero already into FritzGui ,cutechess or any Gui already?
I am running it in xboard via:

xboard -fcp "./lczero --weights get_network -p50000 --noponder" -fUCI and can play games against or let it play against other engines.

Does anybody know how strength increases when the number of playouts is raised?
Playouts play a HUGE role in strenght. All the games I've been running use the default 800, which makes it rather weak but is good for the training purposes, as it provides more games/time. I see you use 50000, that's a nice number and should make it much stronger.

Time control is coming, so playouts will be automatically calculated based on time left / time per move.

Re: LCZero update

Posted: Fri Mar 16, 2018 10:02 pm
by Leo
Werewolf wrote:
Milos wrote:
The similar matches have already been run resulting in strength difference of more than 100Elo.
The whole experiment lacked fairness, which coming from a world-leading organisation is deeply disturbing.
I agree. Its been a while but its always nice to hammer on Google. I am sure their engine has made possible many bad things over the years. They sure as heck host a lot of porn.

Re: LCZero update

Posted: Fri Mar 16, 2018 10:28 pm
by pferd
CMCanavessi wrote:
pferd wrote: Does anybody know how strength increases when the number of playouts is raised?
Playouts play a HUGE role in strenght. All the games I've been running use the default 800, which makes it rather weak but is good for the training purposes, as it provides more games/time. I see you use 50000, that's a nice number and should make it much stronger.
I upped the playouts to 150000 (equals 30-40s/move on my machine) and had a quick match against TSCP 1.81b which was using 1 min/game. The network in use was 1aa02cb1. The game with Leela using the white pieces did not look too bad for :

[pgn]
[Event "Computer Chess Game"]
[Site "brutus"]
[Date "2018.03.16"]
[Round "1"]
[White "lczero"]
[Black "tscp"]
[Result "0-1"]
[TimeControl "60"]
[Annotator "1. -0.04 1... -0.30"]

1.d4
{-0.04/25}
1...d5
{-0.30/6 2.0}
2.Nf3
{-0.10/26 34}
2...Nf6
{-0.25/6 1.9}
3.c3
{-0.12/26 33}
3...Nc6
{+0.02/6 1.9}
4.b4
{-0.34/26 39}
4...e6
{+0.27/6 1.8}
5.a4
{-0.18/26 38}
5...Bd6
{+0.89/5 1.7}
6.Qd2
{-0.01/26 40}
6...O-O
{+1.02/6 1.7}
7.Qb2
{-0.05/26 41}
7...Bd7
{+0.92/6 1.6}
8.b5
{+0.03/26 41}
8...Ne7
{+0.81/6 1.6}
9.Ne5
{-0.04/26 41}
9...Bxe5
{+1.47/6 1.5}
10.dxe5
{-0.14/26 40}
10...Ng4
{+1.48/6 1.5}
11.Na3
{-0.49/26 39}
11...Nxe5
{+1.92/5 1.4}
12.g3
{-0.53/26 34}
12...c5
{+2.06/5 1.4}
13.f4
{+0.05/26 37}
13...Ng4
{+1.94/6 1.3}
14.e3
{+0.23/26 40}
14...e5
{+2.00/5 1.3}
15.Be2
{+0.17/26 40}
15...Qa5
{+1.83/6 1.2}
16.Nb1
{+0.84/26 40}
16...exf4
{+2.06/5 1.2}
17.exf4
{+1.21/26 42}
17...Rfe8
{+1.93/5 1.2}
18.Nd2
{+1.73/26 40}
18...Nf5
{+1.82/5 1.1}
19.Nb3
{+2.56/26 41}
19...Qb6
{+1.87/5 1.1}
20.O-O
{+1.40/26 40}
20...c4+
{+3.67/5 1.1}
21.Nd4
{+1.31/26 42}
21...Nxd4
{+3.67/5 1.0}
22.Bxg4
{-1.92/26 41}
22...Ne2+
{+4.80/5 1.0}
23.Kg2
{-1.64/26 40}
23...Bxg4
{+4.92/5 0.9}
24.Re1
{-1.53/26 40}
24...d4
{+5.55/5 0.9}
25.cxd4
{-2.47/26 40}
25...c3
{+5.59/4 0.9}
26.Qb3
{-2.54/26 39}
26...Qxd4
{+5.96/4 0.9}
27.Rd1
{-3.52/26 40}
27...Qe4+
{+99.93/5 0.4}
28.Kf2
{-28.77/20 2.2}
28...Qf3+
{+99.95/4 0.1}
29.Ke1
{-36.14/19 0.7}
29...Nxc1+
{+99.97/3 0.1}
30.Qe6
{-41.03/18 0.5}
30...Rxe6#
{Xboard adjudication: Checkmate}
0-1
[/pgn]

The reverse game however, was very hard to watch with Leela blundering a knight on move 2.

[pgn][Event "Computer Chess Game"]
[Site "brutus"]
[Date "2018.03.16"]
[Round "2"]
[White "tscp"]
[Black "lczero"]
[Result "1-0"]
[TimeControl "60"]
[Annotator "1. +0.13 1... +0.26"]

1.e4
{+0.13/6}
1...Nc6
{+0.26/25 36}
2.d4
{+0.22/6 1.9}
2...Nxd4
{+0.28/26 37}
3.Qxd4
{+2.30/6 1.9}
3...Nf6
{-0.09/26 31}
4.e5
{+3.00/5 1.8}
4...Ng8
{-0.74/26 38}
5.Nc3
{+2.90/5 1.7}
5...f6
{-0.47/26 36}
6.Bc4
{+2.93/5 1.7}
6...b5
{-0.27/26 38}
7.Bxg8
{+5.41/5 1.6}
7...Bb7
{-1.66/26 38}
8.Bd5
{+5.55/5 1.6}
8...c6
{-1.40/26 38}
9.Bf3
{+5.52/5 1.5}
9...Qb8
{-1.74/26 35}
10.Bf4
{+5.93/5 1.5}
10...d6
{-1.74/26 41}
11.exd6
{+6.05/5 1.4}
11...exd6
{-1.75/26 40}
12.O-O-O
{+6.24/4 1.4}
12...h5
{-2.49/26 39}
13.Qe4+
{+6.73/4 1.3}
13...Kf7
{-1.77/26 42}
14.Nxb5
{+6.82/5 1.3}
14...d5
{-3.61/26 40}
15.Qf5
{+7.32/5 1.2}
15...Qd8
{-3.88/26 40}
16.Nc7
{+7.88/5 1.2}
16...Bc5
{-3.90/26 40}
17.Bxh5+
{+9.43/4 1.2}
17...Kg8
{-7.72/26 40}
18.Qe6+
{+9.59/5 1.1}
18...Kh7
{-7.84/26 41}
19.Bf7
{+9.59/4 1.1}
19...g5
{-6.83/26 38}
20.Qf5+
{+13.27/5 1.0}
20...Kg7
{-8.59/26 39}
21.Qg6+
{+13.32/5 1.0}
21...Kf8
{-12.91/26 34}
22.Ne6+
{+13.37/5 1.0}
22...Ke7
{-13.06/26 33}
23.Nxd8
{+16.17/5 0.9}
23...Bxf2
{-13.17/26 32}
24.Qg7
{+17.10/5 0.9}
24...Be3+
{-13.91/25 36}
25.Bxe3
{+21.65/5 0.9}
25...Raxd8
{-14.22/25 37}
26.Bxd5+
{+23.96/4 0.9}
26...Kd6
{-17.83/25 36}
27.Bg8+
{+23.96/4 0.8}
27...Ke5
{-18.43/25 34}
28.Bd4+
{+25.32/5 0.8}
28...Rxd4
{-17.34/25 33}
29.Re1+
{+99.93/5 0.4}
29...Kf4
{-23.48/24 16}
30.Qxf6+
{+99.95/4 0.1}
30...Kg4
{-26.04/23 1.4}
31.Qf3+
{+99.97/3 0.1}
31...Kh4
{-40.26/18 1.1}
32.Qh3#
{Xboard adjudication: Checkmate}
1-0
[/pgn]

Maybe I was bit too optimistic, but let's see where this will go. Interesting days ahead for us.