It is currently Mon Dec 18, 2017 1:38 pm

All times are UTC - 8 hours [ DST ]




Post new topic Reply to topic  [ 9 posts ] 
Author Message
Offline
 Post subject: alphago self-play reviews
Post #1 Posted: Wed May 31, 2017 9:16 am 
Lives with ko

Posts: 164
Liked others: 14
Was liked: 180
Rank: UK 2d Dec15
KGS: mathmo 4d
IGS: mathmo 4d
I said I'd wait for a few weeks to work on this due to uni exams, but I wonder if I am right that B is already unhappy in game 1 at move 40.

It was quite easy to find key mistakes in the Master vs pro series as it was clear what alphago was punishing, but in the alphago self-play, it seems almost impossible - all you can say is that one side eventually lost, but you don't know where it happened or what they missed.

Though that is hardly the main point of these reviews. The main goal is to explain and describe and add variations.


Alphago's style has developed into playing many light probes in the opponent's area of influence (willing to sacrifice), and tenuki shockingly often (often to play such a probe, before deciding the direction). This is why it is so difficult to review, as the ideas come from a level so far above what we are used to. The material in these games will take many months to be absorbed into the pro go mindset, and I think we will still be finding new things years later. I think it's not unreasonable to say that each game has 10 times the learning material than the master games. Though it's still early days, and we don't know what's yet to come.

I have found that this style of playing sharp but light probes has greatly increased my strength at taking W in handicap games

And there are many brilliant ideas and simple strategies at a level I can appreciate and have applied with success in my own even games.

Here is a quick look at the first 50 moves



Attachments:
1.sgf [6.27 KiB]
Downloaded 1011 times

This post by dhu163 was liked by 5 people: dfan, Gomoto, jeromie, lightvector, Waylon
Top
 Profile  
 
Offline
 Post subject: Re: alphago self-play reviews
Post #2 Posted: Wed Jun 07, 2017 1:21 pm 
Lives with ko

Posts: 164
Liked others: 14
Was liked: 180
Rank: UK 2d Dec15
KGS: mathmo 4d
IGS: mathmo 4d
I'm halfway through 2nd year exams, and I seem to be doing much less revision in exam week anyway. I plan to work harder next year when it is more important and harder, but for this year things are way above expectations given my 4-1 go math ratio, as I'm still on track to scrape a 1st class.

So here's some alphago analysis

Its style is to strive for a particularly high efficiency of every stone it plays. It is much more ready to play light and sacrifice than humans, and the constant tenuki and sharp attachment probes make it so shocking to the eye. But to be 3 stones stronger than top humans, it needs to do shocking stuff.

It is far more ready to fight, attack and risk long dragons in these games. Game 31 was also striking to me for such high quality endgame. But that's probably only as the other games have harder endgame, and but this game's tough endgame at least I can analyse and understand the level of perfection it has reached, B setting up, but W dealing perfectly with a dangerous 2 step ko with the game eventually depending on the final 1/3 point ko, which W won by 2 ko threats, and hence by 1/2 point.

I have emphasised tewari and comparison with local options to tenuki and my usual attempts at broad reasoning behind it's moves

It will truly take years to analyse these games. I picked a random number: game 38 to look at. Even for pros, this will be very difficult. It took them 4 months after the new year's eve Master games to dare to invade 3-3 so early, but these games have 20 times more material to study than the Master games. We need a collaborative effort. Perhaps I even have a bit of advantage over pros with how I have followed the alphago games and thought about it, enough to understand its style quite well.

Just an indepth but very incomplete look at the first 19 moves so far. edit 22:31 : I've added a lot more, but still only up to move 35. It's so difficult



Attachments:
38.sgf [14.34 KiB]
Downloaded 831 times

This post by dhu163 was liked by: dfan
Top
 Profile  
 
Offline
 Post subject: Re: alphago self-play reviews
Post #3 Posted: Wed Jun 07, 2017 3:00 pm 
Lives with ko

Posts: 164
Liked others: 14
Was liked: 180
Rank: UK 2d Dec15
KGS: mathmo 4d
IGS: mathmo 4d


Game 6 seems to have the prettiest shape, it is unbelievable to appear in game: see move 23

moves 36 -70 and particular tesuji at move 47 are also fascinatingly new.


Attachments:
6.sgf [1.97 KiB]
Downloaded 839 times
Top
 Profile  
 
Offline
 Post subject: Re: alphago self-play reviews
Post #4 Posted: Thu Jun 08, 2017 5:00 am 
Lives with ko

Posts: 164
Liked others: 14
Was liked: 180
Rank: UK 2d Dec15
KGS: mathmo 4d
IGS: mathmo 4d
a more in-depth look up at the attachments, up to move 19 in game 1 (edit: now 29) (edit: now quick look at highlights)



again, just up to move 19 game 2, (edit: now quick look at highlights)




this game totally changed how I see sanrensei, and had the most impact on how I played at Challengers', particularly games 5 and 7 (somewhat game 4)

a quick look at game 3 up to move 70


Attachments:
2.sgf [8.45 KiB]
Downloaded 635 times
1.sgf [17.26 KiB]
Downloaded 647 times
3.sgf [6.64 KiB]
Downloaded 635 times


Last edited by dhu163 on Tue Jun 13, 2017 12:02 pm, edited 3 times in total.
Top
 Profile  
 
Offline
 Post subject: Re: alphago self-play reviews
Post #5 Posted: Fri Jun 09, 2017 3:02 am 
Lives with ko

Posts: 164
Liked others: 14
Was liked: 180
Rank: UK 2d Dec15
KGS: mathmo 4d
IGS: mathmo 4d
is it possible to delete posts?
I was hoping just to move this to the post above and delete this to reduce clutter, but I don't know if that is possible.


Last edited by dhu163 on Tue Jun 13, 2017 11:20 am, edited 2 times in total.
Top
 Profile  
 
Offline
 Post subject: Re: alphago self-play reviews
Post #6 Posted: Fri Jun 09, 2017 4:30 am 
Lives with ko

Posts: 164
Liked others: 14
Was liked: 180
Rank: UK 2d Dec15
KGS: mathmo 4d
IGS: mathmo 4d
Its somewhat embarrassing, but I went through game 4, and can't work out why B had to resign (move 244), even though it should be obvious given there aren't many places left to find ko threats.
:scratch:




Attachments:
4.sgf [7.29 KiB]
Downloaded 720 times
Top
 Profile  
 
Offline
 Post subject: Re: alphago self-play reviews
Post #7 Posted: Fri Jun 09, 2017 5:03 am 
Lives with ko

Posts: 164
Liked others: 14
Was liked: 180
Rank: UK 2d Dec15
KGS: mathmo 4d
IGS: mathmo 4d
i think I've solved the problem, but it is surprisingly hard

my end result is W + 1.5 and there may well still be holes in my endgame




Attachments:
4.sgf [8.37 KiB]
Downloaded 719 times
Top
 Profile  
 
Offline
 Post subject: Re: alphago self-play reviews
Post #8 Posted: Thu Jul 20, 2017 10:02 am 
Tengen

Posts: 4325
Location: Cambridge, UK
Liked others: 169
Was liked: 2134
Rank: UK 4 dan
KGS: Uberdude 4d
OGS: Uberdude 7d
Followers of this thread may like to see my review of #38: viewtopic.php?p=221574#p221574

Top
 Profile  
 
Offline
 Post subject: Re: alphago self-play reviews
Post #9 Posted: Sun Oct 15, 2017 10:13 am 
Lives with ko

Posts: 164
Liked others: 14
Was liked: 180
Rank: UK 2d Dec15
KGS: mathmo 4d
IGS: mathmo 4d
my uni maths is going particularly well so far.

as Redmond's reviews turned out a lot better than I expected, I'll look at alphago's games from backwards.

alphago's games are still extremely difficult to review even after 5 months, but I'm pretty pleased with the answers and nice precise sequences I found to answer practically all of my questions in this game up to move 90 ish. I guess the opening is a sort of precise local reading that humans can deal with, where it is mostly a question of finding tesuji and reading forcing lines, without needing to make global judgements.

only move 20 is an unanswered question. move 79 seems questionable, half-inexplicable to use up ko threats (though 5 point reverse sente) unless it is in part to do with bots taking into account the probability of the opponent not responding? Though I do have an explanation locally that I put in the variation, I'm not sure if it sufficient.

especially when alphago plays moves like 37 quite precisely, trying to delay the exchange until it is necessary, or move 80 where two options are almost exactly the same, it does choose the one which is marginally better.



edit: I was struggling with the life and death problem for B on the upper left for a while, and realised that simple looking moves like 104 and 105 have a lot to do with it, so that W 106 gives up on threatening to kill B - it could be played much more severely at B8, but W doesn't think it is killable - B has the wedge at F10 and K4 is probably critical in that fight. Given alphago's reading in game 3, I have to trust that it is right.


Attachments:
49.sgf [11.47 KiB]
Downloaded 280 times

This post by dhu163 was liked by: Shenoute
Top
 Profile  
 
Display posts from previous:  Sort by  
Post new topic Reply to topic  [ 9 posts ] 

All times are UTC - 8 hours [ DST ]


Who is online

Users browsing this forum: Mike Novack and 1 guest


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