Video series: Learning from AlphaGo

gennan · #1

As you may know, DeepMind recently published an opening book annotated with (black's) winrates according to AlphaGo: https://alphagoteach.deepmind.com/

I decided to start a video series about it. This is the first episode:

https://www.youtube.com/watch?v=gGXgu0MMhbg

Fadedsun · #2

I'll check out your video, thanks. I think the Alphago Teach tool is interesting, but wish they had some comments as to why Alpha Go prefers the moves that it does. I guess they probably don't understand any better than we do. As it is, it can give you some ideas on possibly how to play a better opening.

Bill Spight · #3

Fadedsun wrote:

I'll check out your video, thanks. I think the Alphago Teach tool is interesting, but wish they had some comments as to why Alpha Go prefers the moves that it does. I guess they probably don't understand any better than we do. As it is, it can give you some ideas on possibly how to play a better opening.

AlphaGo Teach is a good first effort. We are not going to get much in the way of why AlphaGo thinks that one move is better than others that we can't get elsewhere, because what is new with AlphaGo is what pros did not think of or thought was not so good. The Master series shows that AlphaGo pulled ahead early in most games without its human professional opponents realizing that fact. We are all learning from AlphaGo, and AlphaGo cannot explain its moves.

spook · #4

Thanks for your post gennan, watched it and liked it.
Been studying this position all morning now.

gennan · #5

Episode #2:

https://www.youtube.com/watch?v=ci5AGpUlyP8

djhbrown · #6

Fadedsun wrote:

why Alpha Go prefers the moves that it does.

She doesn't know why she does what she does herself (it just comes out in the wash of statistics), so how could anyone tell you? Mind you, Michael Redmond does a pretty good job of making sense of her plays.

gennan · #7

I agree that Michael Redmond does an excellent job. I highly recommend you all to watch Michael Redmond's AlphaGo games reviews on the AGA youtube channel, if you haven't done so already. They're great!

But besides those videos, I think that we can learn from DeepMind's opening database too, even though it does not contain any explanations.
AlphaGo can't explain why it plays the way it does, but I think the database makes it possible even for amateurs to make educated guesses.
At least that is what I try to do. I share some interesting things that I find in this database and "translate" it into video reviews to make it more accessible for go players less geeky than me.

BTW, I think the content of my videos is most suitable for a player level from about 5k EGF to about 5d EGF.

djhbrown · #8

gennan wrote:

educated guesses

your guess is a lot better than mine, so since you've gone to the trouble of sharing your thoughts, may i ask you about your comment about this position in your first video? Colours and corners are reversed in my picture because that's how Kogo presents them and Sabaki doesn't have a 'switch colours' or 'rotate/flip board' feature.

To get a feel for why either player would want to make any of the few moves they did, one has to see them in context, which is at least the entire forest Kogo sees, which is more than big enough to puzzle this monkey, never mind anything else.

Never mind, plunging into the darkness, i hear you say "Q13 is a decoy stone, to settle P16."

In the Alfie game, there is a white stone on Q4 (ie a black stone on Q16 in your video). Because of this, it looks to me like the meaning of Q13 is to build a position on the right, using P16 as a decoy!

Where am i going wrong?

Could it be that both interpretations are true?, ie white is saying to black "You go one way and i will go the other, and i will get to Scotland before you, 47.5387626889% of the time" (or whatever the number is).

gennan · #9

I reverted to the position in the video after move 4.

Click Here To Show Diagram Code: [go]$$c 1: Black to move. $$ --------------------------------------- $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . O . . . . . . . . . . . . . . . | $$ | . . . , . . b . . , . . . . . X . . . | $$ | . . . a . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . , . . . . . , . . . . . , . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . O . . . . . , . . . . . X . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ ---------------------------------------[/go]

Black now has a choice to focus on the upper side or the left side.

If black wants to focus on the upper side, black could play at G16 at this moment, which could lead to this continuation, which is a semi-joseki (perhaps a bit slack for black, though):

Click Here To Show Diagram Code: [go]$$c 2: Black focusses on the upper side. $$ --------------------------------------- $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . O . . . . . . . . . . . . . . . | $$ | . . . , . . 1 . . 3 . . . . . X . . . | $$ | . . . 2 . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . , . . . . . , . . . . . , . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . O . . . . . , . . . . . X . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ ---------------------------------------[/go]

If black wants to focus on the left side, black usually plays D15 instead of G16. Now it's white's turn. White could focus on taking the corner with a move at C15:

Click Here To Show Diagram Code: [go]$$c 3: White focusses on the corner territory. $$ --------------------------------------- $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . O . . . . . . . . . . . . . . . | $$ | . . . , . . . . . . . . . . . X . . . | $$ | . . 2 1 . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . , . . . . . , . . . . . , . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . O . . . . . , . . . . . X . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ ---------------------------------------[/go]

This white move allows black to switch his focus to the upper side:

Click Here To Show Diagram Code: [go]$$c 4: White focusses on the corner and black focusses on the upper side. $$ --------------------------------------- $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . O 5 . . . . . . . . . . . . . . | $$ | . . 4 3 . . . . . . . . . . . X . . . | $$ | . . 2 1 . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . , . . . . . , . . . . . , . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . O . . . . . , . . . . . X . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ ---------------------------------------[/go]

White can also focus on making potential on the left and / or upper side with a pincer move instead of C15.

If white plays a 2-space high pincer, black plays the large knight move (when D4 is a white stone) and then white attaches at E15, so we arrive at this position (I assume that this is white's plan when he plays the 2-space high pincer):

Click Here To Show Diagram Code: [go]$$c 5: White plays a 2-space high pincer to fight. $$ --------------------------------------- $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . O . . . . . . . . . . . . . . . | $$ | . . . , . . 3 . . , . . . . . X . . . | $$ | . . . 1 4 . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . 2 . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . , . . . . . , . . . . . , . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . O . . . . . , . . . . . X . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ ---------------------------------------[/go]

Now it's black's turn. Now you suggest that black wants to switch focus to the upper side.
With go, the devil is in the details. So my question is: How?

If you see a black continuation from diagram 5 that takes the upper side in a better way than diagram 2, please share it. If it is convincing, you have found a new joseki variation.
So what continuation do you have in mind from diagram 5 to use D15 as a decoy to take the upper side (instead of black E14 to use G16 as a decoy to take the upper left corner)?

Black K16 perhaps?

Click Here To Show Diagram Code: [go]$$c 6: Is this continuation feasible for black? $$ --------------------------------------- $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . O . . . . . . . . . . . . . . . | $$ | . . . , . . 3 . . 5 . . . . . X . . . | $$ | . . . 1 4 . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . 2 . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . , . . . . . , . . . . . , . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . O . . . . . , . . . . . X . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ ---------------------------------------[/go]

When evaluating diagram 6, it's important to consider that it's white's move now and white will probably play a move to attack black's stone at D15. Can white kill it? I would consider white continuations at D16, E16 or D14, but it looks difficult for white to secure the whole upper left as territory.

gennan · **#10**

I tried to query Leela's opinion about Diagram 6:

Click Here To Show Diagram Code: [go]$$c 6: Is this continuation feasible for black? $$ --------------------------------------- $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . O . . . . . . . . . . . . . . . | $$ | . . . , . . 3 . . 5 . . . . . X . . . | $$ | . . . 1 4 . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . 2 . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . , . . . . . , . . . . . , . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . O . . . . . , . . . . . X . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ ---------------------------------------[/go]

It considers D16, E16 and D14, but on my PC it ended up playing F17 (which I didn't consider at all). Then it foresees this continuation:

Click Here To Show Diagram Code: [go]$$c $$ --------------------------------------- $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . 0 . . . . . . . . . . . . . . . . | $$ | . . 9 O . 6 7 . . . . . . . . . . . . | $$ | . . . . . . 3 . . X . . . . . X . . . | $$ | . . . X O 8 . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . O . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . , . . . . . , . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . O . . . . . , . . . . . X . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ ---------------------------------------[/go]

Followed by:

Click Here To Show Diagram Code: [go]$$Bcm11 $$ --------------------------------------- $$ | . . 5 . . . . . . . . . . . . . . . . | $$ | . 3 O 1 . 7 . . . . . . . . . . . . . | $$ | . 6 X O . O X . . . . . . . . . . . . | $$ | . . 4 2 . . X . . X . . . 9 . X . . . | $$ | . . . X O O . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . 8 . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . O . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . , . . . . . , . . . . . 0 . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . O . . . . . , . . . . . X . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ | . . . . . . . . . . . . . . . . . . . | $$ ---------------------------------------[/go]

which Leela considers somewhat favourable for white (55% winrate for white). So Diagram 6 may not be optimal for black, but it probably won't lose you the game (in an amateur game at least). Again, anything is possible (but the devil is in the details).

djhbrown · **#11**

instead of videos or reams of static diagrams, how about turning your comments into an annotated sgf integrated with Kogo's notes (and Redmond's too, where applicable)?

if there were a version of sgf that allowed voice comments instead of text, Go would advance to 20th century technology...

... and if there were a client that matched corner/colour transformations against a game in progress and could one-touch cycle through joseki variation enddpoints, Go education would take a step towards the 21st..

... and if there were one with Swim smarts, it would head towards the 22nd.

you may call me a dreamer, but i am the only one.

PS i dont like black 5 in diagram 6 - seems too slow and not using aji of black 1

gennan · **#12**

djhbrown wrote:

PS i dont like black 5 in diagram 6 - seems too slow and not using aji of black 1

I was just trying to follow through on your original suggestion: focus on the upper side instead of settling in the corner. You did't say how, so I guessed K16 because it takes the upper side. Leela even finds a decent way to use the aji of D15 to reduce white's corner in the process. You think it's not good enough for black, and move 5 at K16 in diagram 6 would be the cause. That's probably true, because D14 is the best move according to pros and AlphaGo. But I assume that your suggestion kind of rules out D14, because AFAIK that would lead to black settling in the corner. So if neither is good enough for you, I suppose you're looking for a way to take the upper side AND deny white territory in the upper left? Sure it would be great for black, but I doubt that it's really possible.

So if you're not satified with that final diagram, I ask you again: where else should black play after white 4 in diagram 6 to take the upper side in a better way than the final diagram from Leela? If you have no execution plan behind your original suggestion, it seems a bit dreamy. Perhaps I'd better stop here trying to finish your dream for you.

gennan · **#13**

djhbrown wrote:

... and if there were a client that matched corner/colour transformations against a game in progress and could one-touch cycle through joseki variation enddpoints, Go education would take a step towards the 21st..

Many joseki have options for both players, so in many instances it's not the case that you can just impose the end result that you prefer. Go is not really as simple as that (like In-Seong says: you're not alone on the go board. Your opponent is there too). And post-joseki follow-up moves and interactions with the rest of the board also matter when comparing different options. So this tool would need to have many annotations / caveats to convey details like that. In the end, the "ultimate joseki dictionary" would have to be almost like AlphaGo to fully inform you on how good the different options are in a particular whole board situation.

djhbrown · **#14**

gennan wrote:

where else should black play after white 4 in diagram 6

Good question - did you ask Leela its "thoughts"? Does the Alfateach sgf provide any variations?

i don't have a definitive answer myself; i just figured that black 3 could have 2 meanings.

Now, after white 4, black needs to reassess the situation. My disliking your black 5 in dia 6 was neither a criticism nor an invitation to debate, it was a student response, a gut feel, made in the hope of learning something as you are stronger than me.

Locally, one generally doesn't ignore a tsuke, so presumably both hanes are worth looking at, with an eye to developing the top or making a position on the left, according to how white responds. I'm sure there are other local moves worth looking at too - for sure, Alfie will have looked at at least 20 choices for white 5.

Being unable to choose myself, i asked Swim. It perceives that D15 is isolated, small, and not cutting off a weak white group, so it doesn't think it should try to save D15 at this stage. So it would go for black moyo expansion + white moyo reduction along/across one of the green lines. Maybe E14 or D13?

D6 seems to have more tension than D8. I haven't asked Swim or Leela about followups.

Attachment:

black5.png [ 31.74 KiB | Viewed 22643 times ]

gennan · **#15**

So how are E14 or D13 contributing to your plan to focus on taking the upper side with black? Note that E14 is the move that pros and AlphaGo use to settle black in the corner and D13 looks like a move to settle black on the left side, both of which are not your intention.

djhbrown · **#16**

gennan wrote:

in many instances it's not the case that you can just impose the end result that you prefer. Go is not really as simple as that

evidently i didnt explain my idea properly - i was just thinking that if a client could match a corner with a joseki dictionary, and could let you flip through the ends of the variations it records in the context of the game, you would have a clearer idea of what initial moves could lead to. i didn't intend to imply that the opponent would dutifully follow any particular line.

my suggestion is based on comments in Kogo, which in several places note general outcomes for different choices (eg corner territory vs outside influence; left vs right, etc).

djhbrown · **#17**

gennan wrote:

So how are E14 or D13 contributing to your plan to focus on taking the upper side with black? Note that E14 is the move that pros and AlphaGo use to settle black in the corner and D13 looks like a move to settle black on the left side, both of which are not your intention.

"my plan"? i never had a plan. my original comment was that it looked to me like white 3 could have two meanings. that's all. if i were white, after black 4, i would try to be flexible and not insist upon a single plan for white 3, which is what i thought i said.

Go is a zero sum game, but i don't think that means that conversations about Go have to be zero sum games too.

Fllecha · **#18**

I want to strongly congratulate with OP to clarify, in two short video, two of my top opening joseki troubles. Many thanks again.

daal · **#19**

I watched the first video in this series and I liked it too. I thought the explanations were clear and made sense, and I also liked that the video was < 15 mins... Looking forward to more!

djhbrown · **#20**

daal wrote:

I thought the explanations were clear and made sense

agree - so much so that they got me thinking and asking a question
At the moment, my understanding of gennan's dia 5 (reproduced below) is: black 3 could have two meanings (i didn't mean to suggest that black should only consider the top right after any white 4).

Black 1 and 3 ask white a question; white's answer at 4 also has two meanings - it either implies she thinks the outside is important, or it's a feint to busy black outside so white can take the inside... so black will reassess things and look for a move to keep her options open, making use of both black 1 and 3.

Attachment:

example4.php.png [ 2.5 KiB | Viewed 22299 times ]

Video series: Learning from AlphaGo

Who is online