Life In 19x19

Dear L19 community,

I'm going to give a talk at my university about the AlphaGo vs Lee Sedol match. The professional An Younggil pointed out how AlphaGo wasted a very strong ko threat in Game 2 against Fan Hui in October; and the general agreement among pros seems to be that, like it is for every go software, ko is a weakness of AlphaGo.

This is why I would like to show an example of a ko fight where it is comparatively easy to show
- how ko threats are made everywhere on the board (the fight jumps here and there)
- and that at some point one of the players will ignore the threat because if it has lower value for him than filling the ko.

The point is that I can then show people move 31 of game 2 of AlphaGo vs Fan Hui and tell them why it is wrong and why I think Lee Sedol will be able to exploit this in march.

Would anyone have a good example of an illustrative ko fight?

Hanspeter

Actually the game that was just played between Murakawa Daisuke 8p and Lee Sedol 9p in the Nongshim Cup is a great example of a game with multiple ko fights, generating and calculating ko threats, etc. A video commentary of the game is available on the AGA channel here (mostly the last hour of the video) and the SGF of the game should show up on go4go pretty shortly.

hanspi wrote:I would like to show an example of a ko fight where it is comparatively easy to show
- how ko threats are made everywhere on the board (the fight jumps here and there)
- and that at some point one of the players will ignore the threat because if it has lower value for him than filling the ko.

I think that that is asking too much of people unfamiliar with the game. If need be, you can just state the second point without proof.

I do think that your audience can understand ko threats that threaten to capture stones immediately. You can show them that easily. Furthermore, in the game threatens to capture some stones.

My suggestion would be to take a final ko fight over a single stone and illustrate it with ko threats that threaten to capture stones. Pro game records hardly ever show such fights, but you can use one of your own games, or make one up.

In addition to an example it is important to mention that winning a ko means that the opponent gets two (or more) moves somewhere else. The effects are nonlocal in time and space. Strong human players will often see that a big ko is possible, and play crazy-looking moves with the main purpose of generating ko threats.

Bill Spight wrote:asking too much of people unfamiliar with the game.

Hi Hanspeter,

I'd like to second what Bill wrote:
even for people familiar with the game, ko can be difficult.
Example: for kyu levels and even low dan levels,
there is still considerable confusion about ko.

Discussion about ko can easily fill a book.

Your target audience has not even digested basic liberties, ataris, and the ko shape yet ( not to mention life-and-death shapes. )
They are raw pedestrians: they're exposed to ataris for the first time in their life only during your lecture.

If you already have a lot of experience interacting with raw beginners, great.
If not, a good preparation for your talk would be to teach complete non-players --
maybe between 25 to 50 individuals, spending at least one hour with each person --
then you find out how much they can absorb, and where they struggle.
This gives you some reference how deep you dig in your upcoming talk.

Dr. Demis Hassabis' lecture is an excellent example of how to talk about Go to a general audience (at Oxford University).

If you really want to illustrate a ko fight to a group of people who are generally unfamiliar with the game, I think you could make up an artificial position that involves a big ko for the life of a group and a number of ataris as ko threats. If you walk through this slowly, I think you could illustrate the importance of a ko fight to someone who knows only the basic rules of the game. As others have said, anything from a professional game is likely to be too subtle to follow. I have trouble following ko fights with the aid of commentary!

But the big problem with talking about ko threats to say black 31 is bad is it is wrong. Black 31 is not bad per se (though it is two moves earlier than the usual timing but there doesn't seem to be a way to turn that into a mistake), the problem is AlphaGo failed to make use of it by playing at a6 for move 35. This is a probe which asks white if he wants to allow black some double sente endgame (downside is the block on lower side is no longer sente) or make the b9 cut tesuji which gives blank two eyes but allows white to take sente. Fan Hui played the a7 hane with perfect timing to turn 31 into a bad move. I would not try to explain this to beginners let alone non-players!

Dear community,

thank you very much for all these useful hints, they have helped me a lot, and have brought me to the realization that I cannot really talk about ko to non-players. I will simply omit this, then.

After all, our talk is more about the AI side and about what might me the problem if AlphaGo just learns by itself from a database of 150000 games. Ko would be one thing, because a drawn-out ko fight is so complex that there may not be enough training data around to really learn it.

So I'll switch over to just using the ladder as an example, which every beginner knows, absolutely needs to know, but needs to know it in order to then NOT PLAY it. In all its game database, AlphaGo has probably seen only a handful of ladders. In the 88000-game GoGoD library I have found only 50 ladder segments, and always in situations where the space was so narrow such that even if the inner player could connect the ladder somewhere, the outer player still had attack potential from it.

So if AlphaGo knows about ladders, it has read about them "between the lines", which is in principle possible for a machine doing deep learning, but it is simple to argue that a teacher showing you a few ladders and then say "don't play this if you are going to lose it" is much more effective.

I can then just mention the ko in the end to tell people that there is a game mechanism in Go that has remote influence just as the ladder has, but with extremely much more complicated dependencies.

Not quite correct, I'm sure, but good enough for non-players.

And some small comments:

@Solomon: I'll look at it for my personal ediucation, thanks!

@Bill Spight: got your point, and it influenced my decision greatly.

@Calvin Clark: true, and this just reminded me that at my present playing strength I should anyway be very reluctant to talk about ko to anyone.

@EdLee: I have very little experience with raw beginners. My 7-year-old (with whom I play 13x13 handicap games) went over the talk wit me, though, and I made sure that he could follow everything. And before you ask: he could follow everything ... except the ko

@jeromie: I thought about this, and when I tried to build one, I realized that I don't think I know as much about ko fights as I should know to construct a simple and good example.

@Uberdude: yes, I see, this does not make it easier, of course. I just saw a wasted ko threat, but not that there could have been another plan which, if followed up, would have been a good plan.

hanspi wrote:So I'll switch over to just using the ladder as an example

Hi Hanspeter,

Interesting you should mention the ladder.
The audience at Dr. Hassabis' speech didn't ask many very Go-specific questions --
but I'd like to ask him if AlphaGo has any "handcrafted" (his term) tweaks at all --
e.g. special ladder-recognition code; Japanese rules vs. Chinese rules, etc. -- Just curious. Thanks.

hanspi wrote:but needs to know it in order to then NOT PLAY it.

Or, play it, depending on the case.

hanspi wrote:My 7-year-old

Congrats. For a great variety of non-players,
the age and other backgrounds matter a great deal.

EdLee wrote:

hanspi wrote:So I'll switch over to just using the ladder as an example

Hi Hanspeter,

Interesting you should mention the ladder.
The audience at Dr. Hassabis' speech didn't ask many very Go-specific questions --
but I'd like to ask him if AlphaGo has any "handcrafted" (his term) tweaks at all --
e.g. special ladder-recognition code;

Yes it does. If you read the Nature paper you will see that ladders are one of the feature planes of the policy neural network (and maybe value too, I can't remember).

Uberdude wrote:

EdLee wrote:

hanspi wrote:So I'll switch over to just using the ladder as an example

Hi Hanspeter,
Interesting you should mention the ladder.
The audience at Dr. Hassabis' speech didn't ask many very Go-specific questions --
but I'd like to ask him if AlphaGo has any "handcrafted" (his term) tweaks at all --
e.g. special ladder-recognition code;

Yes it does. If you read the Nature paper you will see that ladders are one of the feature planes of the policy neural network (and maybe value too, I can't remember).

"Ladders" is nothing AlphaGo could learn from the core material, simply because -- at the level of games used -- you will never find a ladder played out that does not work, nor a ladder that works (with the exception of very short ladders towards the edge of the board).

The same is true for "forbidden moves".

The troubles with embarking on heuristics about ko include the fact that the simple ones are mainly wrong. Unless the ko is "all-dominating", the game gets into a kind of loop, from which either player can exit. Exactly when to exit - well, it's a long story.

Uberdude wrote:Yes it does. If you read the Nature paper you will see that ladders are one of the feature planes of the policy neural network (and maybe value too, I can't remember).

And it is the only one.

The complete list, from the Nature paper: The features that we use come directly from the raw representation of the game rules, indicating the status of each intersection of the Go board: stone colour, liberties (adjacent empty points of stone’s chain), captures, legality, turns since stone was played, and (for the value network only) the current colour to play. In addition, we use one simple tactical feature that computes the outcome of a ladder search.

Cassandra wrote:"Ladders" is nothing AlphaGo could learn from the core material, simply because -- at the level of games used -- you will never find a ladder played out that does not work, nor a ladder that works (with the exception of very short ladders towards the edge of the board).

And this is almost, but not quite, true. I just drew a ladder accross a board and searched for a 5x5 pattern in the GoGoD CD, finding 44 games. I checked a few of them, and all of them were ladders running through the board centre into black-white structures of awesome complexity such that, I think, it was clear to neither player who'd profit more from the ladder.

So without a hand-programmed ladder, AlphaGo would even learn the wrong thing from the database.

I wonder what'd happen if someone took the time to teach AlphaGo the Nihon-Kiin book of Go Proverbs. That contains precisely the sort material AlphaGo wouldn't find in the data. What do you think?

Cassandra wrote:"Ladders" is nothing AlphaGo could learn from the core material, simply because -- at the level of games used -- you will never find a ladder played out that does not work, nor a ladder that works (with the exception of very short ladders towards the edge of the board).

What about the famous game where Lee Sedol played out a long broken ladder, to win the game?