AlphaGo joseki (Zero and Master)

I cataloged a bunch of AlphaGo Zero joseki. I figured it’s nice to observe them all in one place.

AlphaGo Zero joseki:

AlphaGo Master joseki: (from AlphaGo Teaching Tool)

23 Likes

Any comments on why there is no 3-4 joseki from zero v master when by and large this seems to be one of the more common joseki seen in dan and pro play?

1 Like

Yeah, I made a few observations about that:

  1. AG Zero has a strong penchant for 4-4 points. It didn’t play any 3-4 joseki in the Master series because it didn’t play any 3-4 points in that series, only in self-play.

  2. AG Master has a strong penchant for enclosing the corner. AG Zero didn’t play any 3-4 joseki in the Master series because AG Master enclosed all the 3-4 points by move 5.

Because of the low sample of 3-4 and 3-3 joseki, I used the 20 blocks series to “flesh out” those joseki sequences.

Interesting. Especially when you compare the win percentages on something like waltheri between ni-ren-sei openings vs the chinese variants.

And as white AlphaGo Zero systematically plays two 4-4 points.

Nice indeed to have those joseki in one place.

1 Like

In the end, I guess, starting at tengen isn’t efficient enough. None of them have done it.

3 Likes

Couldn’t it be that AGZ hasn’t yet advanced to that point in its self-learning? That playing a billion more games would perhaps show that Tengen, in fact, is The Move?

3 Likes

Then, we should just wait and see. For now, you can beat 9p if you start on 4-4.

4 Likes

:joy: good one!

As I understand it, AlphaGo improves by eliminating bad moves rather than finding good ones. If tengen was leading to winning games, it would still be an option.

We are not seeing any 3-5 or 4-5 opening either btw.

1 Like

I don’t think that’s strictly true. It might depend on what the difference is between “eliminating bad moves” and “finding good ones” is.

It basically explores possibilities, randomly.

So I actually wonder whether it got around to trying out tengen much. Tengen is special, but it wouldn’t “know” that unless it actually tried it. And it is only one amongst 361 points to try. It is 4 times more likely to try any other point that tengen, due to symmetry.

And once tengen falls behind in being tried, I think its conceivable that it gets a bias to be not tried, though this is just an intuition on my part.

I was keenly interested to see if it would develop tengen openings, because I remember reading somewhere an observation that pros don’t do it mostly because the possibilities are too expansive from there: they prefer corner based openings as much because they require less reading than any other reason. This might not apply to AlphaGo, so it would have been nice to see it make something of it :slight_smile:

1 Like

Added 4-4 enclosure joseki. See top post.

3 Likes

Added 4-4 double approach joseki. See top post.

This is probably the final update, unless someone recommends another pattern to look at.

3 Likes

With Deepmind’s release of the AlphaGo Teaching Tool and handy-dandy SGF (which is pattern-searchable in the Kombilo database program), I made a puzzle showcasing some top joseki patterns by AlphaGo Master.

There’s too many patterns to describe even in several puzzles, so I restricted joseki patterns to moves that were played at least 10% of the time in each position and appeared more than 2 times in the SGF. I also added some evaluation scores based on a quick and dirty visual of the Teaching Tool. Enjoy!

6 Likes

It’s interesting that in response to a pincer on a small knight approach to 4-4, leaping in to the 3-3 is “merely situational”, not joseki,

Interesting because Dwyrin teaches that this is pretty much what you should do unless there’s some situation demanding otherwise…

1 Like

One thing AlphaGo taught us is humility. Currently there is no such thing as “what you should do” in the fuseki. AlphaGo has left 9Ps puzzled, don’t expect amateurs to do any better.
One of my favorite series is Michael Redmond’s commentary on the AlphaGo games, on the Youtube AGA channel. A 9 pro who is not afraid to say “I don’t know”.

Chess players are in the same boat. I was watching a commentary of an Alpha Zero game, and the commenter was puzzled to see Alpha move its queen over and over again in the opening.

Interesting indeed. Conventional wisdom was that the best approach to 4-4 was a small knight move, and leaping into the 3-3 was a natural follow-up. Maybe AlphaGo thinks that if you are going for the 3-3 invasion, you’d better do it right away and just skip the small knight approach…

2 Likes

Wow yeah - I’ll have to start looking at the board that way and see if it makes sense.

A counter suggestion (I mean the answer to the question why didn’t I just invade instead of approach) is that “I didn’t want to invade 3-3, I wanted a peaceful joseki, but my darn opponent pincered”.

This is a fascinating topic. So I’m going to put some thoughts out there, even if it’s a little long :slight_smile:

The reason I find it fascinating is looking at the struggle to work out what to do with this knew knowledge - the knowledge that what our Dans know is not “the best there is”.

One reaction to this new knowledge is “therefore there is nothing we know about what you should do anymore”.

I actually think this “there is no such thing as what you should do” actually should only be directly applied to Pros aspiring to beat AlphaGo. They need their minds completely open to the new exploration.

But the new experience from AlphaGo doesn’t negate the years and depth of experience of human play about ho to learn the game and how to learn to progress up from the basics to Dan.

It in fact just tells us that there’s another layer to master (or many more) than we had thought before.

For example, we used to joke (before AlphaGo) “Dans are crazy, don’t try to understand them”.

We did not say “because we can’t understand what Dan’s are doing, there is no ‘what you should do’”.

In fact, we understood that there are layers of learning. There is the initial learning, that TPKs do: “corner side centre” etc. Then there is the learning we do to move up through DDK and strive for SDK. This is very solid learning, and I assert that it is not undone by AlphaGo.

Then there was SDK and finally crazy Dan level play. The Dans used to teach us how to progress through these layers from a platform of certainty about their own knowledge.

What AlphaGo did is challenge their knowledge, and make them realise that they too don’t have just incremental learning to progress up through Dan, but a whole massive unexplored field of how to play.

And for sure as they discover things out there, new knowledge will flow down, be “captured and simplified”, for us lower people to use. One such example might be the overthrow of “don’t invade 3-3 early”. But this doesn’t throw out all the old guidlines and knowledge, just modifies and extends it.

Therefore, I think it is not right to answer any question about “what should we do” with “there is no more should”.

It’s just like before: to answer that question, you need to know what level the asker is asking from and give them the suitable guidance, based on the wealth of experience we humans draw on and have systematised into strategies and guidelines (proverbs).

GaJ

  • TPK: twenty plus KYU. I deliberately call this layer out, because it is easy to lump them with DDK, but 19k is like a god to a 23k: the challenges that DDKs face are diferent to those of TPK learners. I think the lack of this distinction does the Go community a disservice, because beginners can’t tell what advice is stuiable for them, compared to advice suitable for people in the low DDKs.
1 Like

It’s really interesting to me, too. AG Zero’s ideal 3-3 pattern (which AG Master also rates highly) is identical to AG Master’s ideal small-knight, pincer, 3-3 pattern. The only difference is the outer two stones (small knight and pincer), an exchange which doesn’t appear to matter very much. My take away is that 3-3 is still a joseki of sorts, but AG Master doesn’t rate the pincer highly to begin with, and it thinks other moves better “punish” the pincer.

Keep in mind that when Dwyrin says 3-3 is generally what you should do after the pincer, he’s thinking about the traditional joseki. Master apparently agrees with him that the traditional joseki good for the 3-3 player. The big game-changer is that AG Master thinks both players made mistakes to reach the traditional result. Specifically, the pincer player’s connection (last move of the joseki) is a mistake. AG Master prefers a second line push, after which it likes the joseki for the pincer player. Again, really interesting.

2 Likes

3-4 joseki is now complete. See top post.

1 Like