AlphaGo Zero Patterns
-
Uberdude
- Judan
- Posts: 6727
- Joined: Thu Nov 24, 2011 11:35 am
- Rank: UK 4 dan
- GD Posts: 0
- KGS: Uberdude 4d
- OGS: Uberdude 7d
- Location: Cambridge, UK
- Has thanked: 436 times
- Been thanked: 3718 times
Re: AlphaGo Zero Patterns
Nice article, but I don't think not playing the wedge in double approach is a blind spot for humans: I've had a general preference for corner hane for some years now and that's based on my understanding that plenty of human pros (particularly Chinese?) have preferred it since quite some time before AlphaGo existed.
Also 7 in pattern 1 rather than direct capture surprised me. Does it fear Black tenuki if capture so loses a little extra points to be sure of sente? I suppose p14 is also a possible shape followup and it creates more ko threats.
Also 7 in pattern 1 rather than direct capture surprised me. Does it fear Black tenuki if capture so loses a little extra points to be sure of sente? I suppose p14 is also a possible shape followup and it creates more ko threats.
-
RobertJasiek
- Judan
- Posts: 6273
- Joined: Tue Apr 27, 2010 8:54 pm
- GD Posts: 0
- Been thanked: 797 times
- Contact:
Re: AlphaGo Zero Patterns
I am only surprised by one move:
At first, I thought that the following provides better eye shape and better endgame:
But maybe AlphaGo Zero jugdes that this later attack is more relevant:
EDITS
At first, I thought that the following provides better eye shape and better endgame:
But maybe AlphaGo Zero jugdes that this later attack is more relevant:
EDITS
-
Uberdude
- Judan
- Posts: 6727
- Joined: Thu Nov 24, 2011 11:35 am
- Rank: UK 4 dan
- GD Posts: 0
- KGS: Uberdude 4d
- OGS: Uberdude 7d
- Location: Cambridge, UK
- Has thanked: 436 times
- Been thanked: 3718 times
Re: AlphaGo Zero Patterns
Robert, for interest, there was a somewhat similar shape in AG Master self-play #38 (see viewtopic.php?p=221574#p221574) and in that case it connected rather than capture on move 33, I think because it didn't like the 2-1 throw-in ko.
-
sorin
- Lives in gote
- Posts: 389
- Joined: Wed Apr 21, 2010 9:14 pm
- Has thanked: 418 times
- Been thanked: 198 times
Re: AlphaGo Zero Patterns
Indeed, the popularity of the wedge vs hane declined in time, but it still appears in pro games even in 2016. AlphaGo seems to a have much stronger opinion about itUberdude wrote:Nice article, but I don't think not playing the wedge in double approach is a blind spot for humans: I've had a general preference for corner hane for some years now and that's based on my understanding that plenty of human pros (particularly Chinese?) have preferred it since quite some time before AlphaGo existed.
Maybe it likes that it gets some forcing moves from the outside - like P15 - or just in general the fact that black ends up with one extra group (in theory)?Uberdude wrote: Also 7 in pattern 1 rather than direct capture surprised me. Does it fear Black tenuki if capture so loses a little extra points to be sure of sente? I suppose p14 is also a possible shape followup and it creates more ko threats.
(By the way, moves 5 and 7 are interchangeable - in one of the games it plays at 7 first).
I looked this up in my SGF collection and found only one human game with a similar starting pattern, except that white's approach move at R12 is missing (there is one two spaces down instead); the one "in the corner" is some Inoue Ichiro, while "outside" is Go Seigen, who simply captured the stone like you suggested.
-
sorin
- Lives in gote
- Posts: 389
- Joined: Wed Apr 21, 2010 9:14 pm
- Has thanked: 418 times
- Been thanked: 198 times
Re: AlphaGo Zero Patterns
This last diagram looks very convincing for why the capture is better!RobertJasiek wrote: But maybe AlphaGo Zero jugdes that this later attack is more relevant:
-
Bill Spight
- Honinbo
- Posts: 10905
- Joined: Wed Apr 21, 2010 1:24 pm
- Has thanked: 3651 times
- Been thanked: 3373 times
Re: AlphaGo Zero Patterns
I suspect that AlphaGo Zero has a path dependency. That is, even though it trains against a previous version, and ought to correct errors eventually, one version and the next share many assumptions, not all of which are correct.sorin wrote:Indeed, the popularity of the wedge vs hane declined in time, but it still appears in pro games even in 2016. AlphaGo seems to a have much stronger opinion about itUberdude wrote:Nice article, but I don't think not playing the wedge in double approach is a blind spot for humans: I've had a general preference for corner hane for some years now and that's based on my understanding that plenty of human pros (particularly Chinese?) have preferred it since quite some time before AlphaGo existed.
The main focus of the AlphaGo team is machine learning, not go learning. That is, their aim was to produce a strong go program, not to answer questions about go. It would be quite interesting to see AlphaGo Zero.Two, an independent program which follows its own path. They might well play different styles and favor different plays. My guess is that they would.
The Adkins Principle:
At some point, doesn't thinking have to go on?
— Winona Adkins
Visualize whirled peas.
Everything with love. Stay safe.
At some point, doesn't thinking have to go on?
— Winona Adkins
Visualize whirled peas.
Everything with love. Stay safe.
-
sorin
- Lives in gote
- Posts: 389
- Joined: Wed Apr 21, 2010 9:14 pm
- Has thanked: 418 times
- Been thanked: 198 times
Re: AlphaGo Zero Patterns
That is a fascinating topic, it occurred to me too - the question "would the learning from scratch always converge to the same point/style".Bill Spight wrote: I suspect that AlphaGo Zero has a path dependency. That is, even though it trains against a previous version, and ought to correct errors eventually, one version and the next share many assumptions, not all of which are correct.
The main focus of the AlphaGo team is machine learning, not go learning. That is, their aim was to produce a strong go program, not to answer questions about go. It would be quite interesting to see AlphaGo Zero.Two, an independent program which follows its own path. They might well play different styles and favor different plays. My guess is that they would.
Given that AlphaGo Zero ended up playing pretty similarly to the human style though (favoring after all 3rd and 4th lines mostly, in the beginning of the game) in both of 20-block and 40-block versions, seems to suggest that there is "style convergence" (to some extent).
I would like very much to see the "trajectory" it took during learning; the published paper showed the frequency of a few patterns, but I would like to see in more detail how does the style change between amateur 1-dan and pro 9-dan stages...
-
moha
- Lives in gote
- Posts: 311
- Joined: Wed May 31, 2017 6:49 am
- Rank: 2d
- GD Posts: 0
- Been thanked: 45 times
Re: AlphaGo Zero Patterns
I think it's not just coincidence that AG0 ended up at very similar levels than AGM, only a little bit stronger. (And there may be other reasons besides the trivial one: DM only developed it until it was able to reach this goal.)sorin wrote:That is a fascinating topic, it occurred to me too - the question "would the learning from scratch always converge to the same point/style".Bill Spight wrote:It would be quite interesting to see AlphaGo Zero.Two, an independent program which follows its own path. They might well play different styles and favor different plays. My guess is that they would.
Given that AlphaGo Zero ended up playing pretty similarly to the human style though (favoring after all 3rd and 4th lines mostly, in the beginning of the game) in both of 20-block and 40-block versions, seems to suggest that there is "style convergence" (to some extent).