It is currently Wed May 22, 2019 12:03 pm

All times are UTC - 8 hours [ DST ]




Post new topic Reply to topic  [ 3 posts ] 
Author Message
Offline
 Post subject: Is it so hard for LZ to understand this?
Post #1 Posted: Mon Feb 11, 2019 6:04 pm 
Dies with sente

Posts: 85
Liked others: 2
Was liked: 12
Rank: KGS 2 D
In this 2 handi game with Phoenixgo, Leela Zero #203 got winrate as high as 90%, but failed to recognize the danger of filling own liberty. I tried to increase the playout up to 10k and still LZ was blind about it (white plays A to atari and black answers, then white must prevent black to play B to make use of the dead stone. However, LZ can't see this)

Attachment:
LZ-filling-own-lib.png
LZ-filling-own-lib.png [ 507.25 KiB | Viewed 553 times ]

Attachment:
203vc-phoenix-20s-b.sgf [3.52 KiB]
Downloaded 29 times

Top
 Profile  
 
Offline
 Post subject: Re: Is it so hard for LZ to understand this?
Post #2 Posted: Mon Feb 11, 2019 8:10 pm 
Honinbo

Posts: 8379
Liked others: 2432
Was liked: 2905
In playing around with Leela 11 (Deep Leela) I have seen worse. All I can think of is that, despite their whole board skill, today's top bots have no particular local reading skill, no go knowledge, and no logic per se. So occasionally they may run across a type of urgent local situation which they have not encountered before in their training or which, if they have encountered it, they misplayed with self play, and therefore are unaware of the urgency and have a blind spot. In infinite time they can be trained to recognize the urgency and will read the local situation deeply enough to make the right play, but that has not happened for this situation yet.

As for why LZ has trouble with filling its own liberty in general, I think that Bill Taylor inadvertently supplied the answer over 20 years ago on rec.games.go. He suggested evaluating all plays as gote. OC, that gives different answers from standard evaluation. It also evaluations sente plays more highly than standard evaluation, and so recommends that they be played early. In this case filling the dame gives White a large threat. LZ has no go knowledge, and so knows nothing about sente. So in effect it evaluates plays as though they were gote. White can see that if Black does not save her ataried stones White's estimated win rate will go up significantly and so plays the atari, filling its own liberty. Without locally deep reading LZ will see no danger. A human with knowledge of sente can see that as a sente the atari is worth at most a point or two and will not be tempted to play it at this time.(OC, the human could still blunder later, but that's another question.)

_________________
There is one human race.
----------------------------------------------------

The Adkins Principle:

At some point, doesn't thinking have to go on?

— Winona Adkins

Top
 Profile  
 
 Post subject: Re: Is it so hard for LZ to understand this?
Post #3 Posted: Mon Feb 11, 2019 11:06 pm 
i can't comment on gote or sente, not here, anyway :) - but i think i can explain why LZ has this bizarre behaviour, which turns up in LZ local life and death playouts, which irgobot has to oversee to prevent LZ from killing herself.
irgotbot wrote:
Laizy is such a silly girl!
it boils down to this: LZ knows nothing about Go except, ironically, the no-suicide rule, and how to count the score at the end of a game with no more legal moves - all her thinking is guesswork, driven by her probability calculations backed up through her policy/value net, trained over thousands of years of virtual human time in a few minutes (hours, anyway) of actual computer time by god only knows how many pcs pushing on the cloud.

Because LZ knows nothing, she will always play a move if there's anywhere she can play one. For very good reasons, she has never heard of Pass - because during training, when one side is miles ahead, Pass can still produce a win% high enough to be upvoted, and you would end up with a bot that looked altogether supercilious, as she repeatedly Passes while you are desperately trying to save your bacon, which you shouldnt do, despite what modern teachers say, because it's bloody annoying and disrespectful to the opponent when you carry on when all was already lost ages ago.

LZ (and all the rest of Alphago's children) is the worst offender ever at never resigning; that's why the Alphago team had to do a Dave* on her when Lee Sedol found a flaw in her armoury way back when.

So what does all this mean? It means that LZ cannot be trusted to do the decent thing, so it's up to the client to stomp on her when she should shut up and bow out, or dance around the room with her hands in the air if she's won hands-down. As yet, as far as i know, irgobot is the only client with the temerity to tell LZ to not fill in her own eyes - but if you are patient, others will pick up the reins too, eventually.

PS LZ can see the sequence you mentioned, but she doesn't care about it! A win is a win, and any win is good enough, so she chooses the kind of move she has been trained to most favour by her wierd (sic) and wonderful "deep" learning.

*
HAL wrote:
Daisy, Daisy, give me the answer, do...

Top
  
 
Display posts from previous:  Sort by  
Post new topic Reply to topic  [ 3 posts ] 

All times are UTC - 8 hours [ DST ]


Who is online

Users browsing this forum: cobrahc, Google [Bot] and 3 guests


You cannot post new topics in this forum
You cannot reply to topics in this forum
You cannot edit your posts in this forum
You cannot delete your posts in this forum
You cannot post attachments in this forum

Search for:
Jump to:  
Powered by phpBB © 2000, 2002, 2005, 2007 phpBB Group