Life In 19x19
http://lifein19x19.com/

DeepZenGo thread
http://lifein19x19.com/viewtopic.php?f=15&t=14576
Page 2 of 4

Author:  Shenoute [ Sun Oct 22, 2017 8:31 am ]
Post subject:  Re: DeepZenGo thread

I've had a look at DEEPZEN(B)'s account on Tygem, it currently stands at 2598 wins and 632 losses. It played a lot until July 27th and was losing several games each day (4W-4L on the 27th, 16W-4L on the 26th for instance). After July 27th, it stopped playing and only came back the day before yesterday, October 20th. Since then it has lost only two games (71W-2L).

As these games do not seem to be available anywhere except Tygem and cannot be downloaded there, I replayed them and created sgf files of the two losses. I haven't looked at them in detail yet but something weird seems to have taken place in the game against TuTu (I entered the result as given on Tygem).





Attachments:
DEEPZEN(B)-TuTu(P)_20171021 (Tygem).sgf [2.46 KiB]
Downloaded 587 times
tmch-DEEPZEN(B)_20171022 (Tygem).sgf [1.31 KiB]
Downloaded 593 times

Author:  Uberdude [ Sun Oct 22, 2017 10:05 am ]
Post subject:  Re: DeepZenGo thread

I believe tcmh is Tong MengCheng 6p, one of the top Chinese pros just behind the top tier of world champions, but can beat them on a good day; currently goratings #32. First game looks like Zen misunderstood the dead big eye? That's the kind of thing that's very rare to actually happen in a real game, but a 20 kyu learns in a book. So maybe a gap in its knowledge (and the large number of external liberties means the tree search doesn't kill it). I wonder how AlphaGo Zero deals/learnt with such things.

Author:  moha [ Sun Oct 22, 2017 11:19 am ]
Post subject:  Re: DeepZenGo thread

Uberdude wrote:
First game looks like Zen misunderstood the dead big eye? That's the kind of thing that's very rare to actually happen in a real game, but a 20 kyu learns in a book. So maybe a gap in its knowledge (and the large number of external liberties means the tree search doesn't kill it). I wonder how AlphaGo Zero deals/learnt with such things.
I think dead shapes are to be recognised by the net(s), the search wouldn't normally kill/capture them anyway since that would not be a good line. You may be right with gap in knowledge in this case, but on second thought the earlier corner placement doesn't seem like something the net could miss or rank too low, so maybe both moves was searched but not deep enough to understand the ko? Or there may also be a bug somewhere.

About AG it is pure speculation, but I think it is more search oriented or at least has much deeper and more accurate searches (esp. zero with its huge speedup). Also the randomized training of zero probably covers rare shapes like this as well. And of course, even AG has shown a tactical artifact once. :)

Author:  Bill Spight [ Sun Oct 22, 2017 11:35 am ]
Post subject:  Re: DeepZenGo thread

It is important to remember that the current crop of strong go programs do not rely much upon logic. Logic only comes into the picture when they can read the game out, which they cannot, until maybe near the end of play. Their search is probably good enough so that they could read out difficult life and death positions locally, but they do not use local search, it's all global search. For instance, human beginners can learn basic ladders because they can be deduced by logic, but it took AlphaGo Zero a long time to learn basic ladders, and it probably does not know them perfectly, not to mention the complex ladders of some problems.

This kind of example makes me more inclined to think that a team of human plus super-human program may well be better than the super-human program alone. :)

Author:  pookpooi [ Sun Oct 22, 2017 11:39 am ]
Post subject:  Re: DeepZenGo thread

Bill Spight wrote:
This kind of example makes me more inclined to think that a team of human plus super-human program may well be better than the super-human program alone. :)

I heard that it goes opposite way in chess, but chess and go are completely different things.
But everyone already knows about DeepZen vs. Kong Jie + CGI match, we need more games like this though before conclude anything. Plus CGI is not on the super-human level yet (it's playing on Tygem right now and is only as strong as Leela)

Author:  moha [ Sun Oct 22, 2017 11:50 am ]
Post subject:  Re: DeepZenGo thread

pookpooi wrote:
Bill Spight wrote:
This kind of example makes me more inclined to think that a team of human plus super-human program may well be better than the super-human program alone. :)
I heard that it goes opposite way in chess, but chess and go are completely different things.
But everyone already knows about DeepZen vs. Kong Jie + CGI match
Actually I'm interested in the same thing but with a slight difference: instead of a human using an AI as assistance (for tactical verification or as an idea source etc.), the other way may prove to be more effective.

So a human pro would take the evaluation of the AI, and usually let it through. Only if the pick has a bad feeling in it from a human strategic point, going against principles, strange choice of direction etc. (basically, a potential horizont effect) AND the second or third pick of the AI is not much lower in winrate but looks more healthy, would the human intervene. I would be very surprised if this kind of filtering would not raise the strength - but at least it should not lower it. :)

Author:  Bill Spight [ Sun Oct 22, 2017 12:04 pm ]
Post subject:  Re: DeepZenGo thread

pookpooi wrote:
Bill Spight wrote:
This kind of example makes me more inclined to think that a team of human plus super-human program may well be better than the super-human program alone. :)

I heard that it goes opposite way in chess, but chess and go are completely different things.


Well, what I heard about chess was different. Maybe how the team works is different. I mean that the human considers the top plays suggested by the program, along with their evaluations, and decides on the actual play, which may not be one of the program's suggestions. By contrast, a team where the program makes some plays and the human makes others would almost surely be weaker than the machine alone. Anyway, the human-machine team could always be as good as the machine alone if the human always chooses the machine's favored play. ;)

Author:  Uberdude [ Mon Oct 23, 2017 2:36 am ]
Post subject:  Re: DeepZenGo thread

Zen is also pretty good at 5-in-a-row :lol:

Click Here To Show Diagram Code
[go]$$W
$$ | . . . , . . . . . , .
$$ | . . X . . . . . . . .
$$ | . . . . . . . . . . .
$$ | . 4 . . . . . . . . .
$$ | . 3 X 2 . . . . . . .
$$ | . . 1 . . . . . . . .
$$ | . . . O . X . . . , .
$$ | . . . . 5 6 . . . . .
$$ | . . . . . 7 8 . . . .
$$ | . . . . . . . . . . .
$$ +----------------------[/go]

Author:  pookpooi [ Mon Oct 23, 2017 2:43 am ]
Post subject:  Re: DeepZenGo thread

Now can DeepZen take on GodMoves when (s)he play seriously? #AskingTheRealQuestion #RevengeMatch #AnonymousPlayer

Author:  Shenoute [ Mon Oct 23, 2017 4:22 am ]
Post subject:  Re: DeepZenGo thread

Team human scored 4 wins today (for 20 losses), a marked improvement over the previous days. One of these however was on two stones, making it maybe less surprising but no less interesting.



The other three games are equally interesting I feel. The one against 5353 features large exchanges on the right side in the beginning and two simultaneous ko in the end where only Deepzen played threats.



In the game against KeA, Deepzen seems to have misread(?) something in the upper right (move 138!) and also the semeai at the end.



Attachments:
DEEPZEN(B)-CHAH(P)_20171023 (Tygem).sgf [1.95 KiB]
Downloaded 555 times
DEEPZEN(B)-5353(P)_20171023 (Tygem).sgf [2.02 KiB]
Downloaded 602 times
DEEPZEN(B)-KeA(P)_ 20171023 (Tygem).sgf [2.08 KiB]
Downloaded 532 times

Author:  Shenoute [ Mon Oct 23, 2017 4:23 am ]
Post subject:  Re: DeepZenGo thread

Finally the fourth loss (see the post above for the other three) shows that Deepzen still has problems with dead shapes.



Attachments:
Mollayo(P)-DEEPZEN(B)_20171023 (Tygem).sgf [1.44 KiB]
Downloaded 558 times

Author:  moha [ Mon Oct 23, 2017 6:31 am ]
Post subject:  Re: DeepZenGo thread

Shenoute wrote:
Finally the fourth loss (see the post above for the other three) shows that Deepzen still has problems with dead shapes.
I think / agree with Uberdude that it is more surprising that AG doesn't have this kind of problem than DeepZen having it. If I heard about the neural search idea for the first time today, my first thought would be that won't it miss key moves sometimes? (In the Sada Atsushi position, it probably didn't miss the corner placement, but maybe missed one of the responses after the hane, where the opponent sets up the ko.) So what can AG do differently?

1. More processing power, more search. This is not a good explanation in itself, since having even 100 times more performance is still negligible for wider searches. I think a program can only hope to search about 2-3, max 4 candidates on average.

2. Better net. This is at least a possibility, but I still don't think it is realistic for a net never to miss key moves (or rank them just a bit too low).

3. Better connection between the net and the search, better search techniques (such as when to widen the search, when to deepen it etc.) This may be a possibility, but nothing tangible at this point.

I think the most likely explanation comes from 1+3: maybe AG does significantly wider searches around the top of the tree (first 2-3 moves or so), and only prune heavily deeper. This would explain the lack of visible oversights, but would still leave it with unnoticed weeknesses where key moves are missed deeper in the tree, leading to suboptimal choices at top.

Author:  Shenoute [ Mon Oct 23, 2017 7:52 am ]
Post subject:  Re: DeepZenGo thread

moha wrote:
(...)
I think the most likely explanation comes from 1+3: maybe AG does significantly wider searches around the top of the tree (first 2-3 moves or so), and only prune heavily deeper. This would explain the lack of visible oversights, but would still leave it with unnoticed weeknesses where key moves are missed deeper in the tree, leading to suboptimal choices at top.

Yes, you're probably right. AI does work in mysterious ways :-)

Author:  Shenoute [ Mon Oct 23, 2017 8:15 am ]
Post subject:  Re: DeepZenGo thread

Bad day for Deepzen, it just lost against hanuu a Japanese 9p on Wbaduk. I've watched a couple dozen of Deepzen's games there and it is the first I see him lose (it seems that it is not possible anymore to view players' past games with WBaduk client so it's harder to keep track of how he's doing).

I 've got the feeling that deepzengo was outplayed very early at the bottom and could not catch up...



Attachments:
deepzengo-hanuu_201723 (WBaduk).sgf [1.39 KiB]
Downloaded 532 times

Author:  Uberdude [ Mon Oct 23, 2017 10:35 am ]
Post subject:  Re: DeepZenGo thread

Followers of the games on Tygem might find this list of usernames and the pros behind useful e.g. DeepZenGo just beat HopeIdo aka Meng Tailing 6p.

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

Author:  pookpooi [ Mon Oct 23, 2017 10:44 am ]
Post subject:  Re: DeepZenGo thread

Is it become accepted that AlphaGo can handle Japanese rule and its complication? Because I thought that at first but some Reddit comment said to me that it can't handle yet, AlphaGo just so strong it doesn't matter anymore. So I'm a little bit confused here. Applaud DeepZen that it can, mostly handle Japanese rule.

Author:  Uberdude [ Tue Oct 24, 2017 4:13 am ]
Post subject:  Re: DeepZenGo thread

Some interesting fighting between big sticks of stones in Zen vs Tsurata Kazushi 4p whose kick of move 6 was perhaps AlphaGo Zero inspired:


Author:  Shenoute [ Tue Oct 24, 2017 5:39 am ]
Post subject:  Re: DeepZenGo thread

Uberdude wrote:
Some interesting fighting between big sticks of stones in Zen vs Tsurata Kazushi 4p whose kick of move 6 was perhaps AlphaGo Zero inspired:

Nice. Around move 56, I think my group would die regardless of which color I would play :-)

Author:  Shenoute [ Wed Oct 25, 2017 1:15 am ]
Post subject:  Re: DeepZenGo thread

Zen went 26W-3L on the 24th and is currently at 19W-1L as far as the 25th is concerned.

Something strange just happened in its last game. Zen played a double ko he could not win (in the bottom right corner). In one of the games I posted earlier in this thread (the one against 5353), Zen also tried to play two ko simultaneously.


Here is a very short game (111 moves) won today by XIUZHI(P).


And to make good measure, here is a even shorter win (94 moves) yesterday by Zen, a nice kill in the center.


Attachments:
DEEPZEN(B)-duodi(P)_20171024 (Tygem).sgf [776 Bytes]
Downloaded 420 times
DEEPZEN(B)-XIUZHI(P)_20171025.sgf [896 Bytes]
Downloaded 735 times
DEEPZEN(B)(9D)_Mollayo(P)(9D)_201710250925.sgf [2.01 KiB]
Downloaded 447 times

Author:  Schachus [ Wed Oct 25, 2017 2:53 am ]
Post subject:  Re: DeepZenGo thread

Shenoute wrote:
Zen went 26W-3L on the 24th and is currently at 19W-1L as far as the 25th is concerned.
Here is a very short game (111 moves) won today by XIUZHI(P).



that would be Park Jungwhan, right?(I found this name from the senseis page about the masters games, where he he also played a couple of games)

So it seems Zen is still not over the top of the very top professionals

Page 2 of 4 All times are UTC - 8 hours [ DST ]
Powered by phpBB © 2000, 2002, 2005, 2007 phpBB Group
http://www.phpbb.com/