Thread: "AlphaGo, 4D Go, Hyperbolic Go"

--001a11408ca055e6d4052da12bd4
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: quoted-printable

Anyone catch the match last night? Melinda and I did, and are
enthusiastically discussing it. It was awesome! You can watch the
remaining games live here
.

To connect the excitement back to the group, I wanted to mention you can
play Go on the 1-skeletons of 4D polytopes using an early version of
Jenn3D. Head to the very bottom of this page
to try. Duoprisms are particularly
interesting, because you can use them to make boards that remove all the
edges of a traditional board but are otherwise the same. Playing on
polytopes feels like it would generally have too much freedom though,
especially if single stones have more than 4 adjacent liberties.

Adapting MagicTile to support Go might work well, since it would keep the
boards as 2D surfaces. A {5,4} tiling would be a natural choice for a
first board, and probably some of Andrea Hawksley's ideas about
non-euclidean
chess would apply. But I also wonder if hyperbolic Go would be
fundamentally flawed. Random walks in the Poincar=C3=A9 disk inevitably es=
cape
to infinity. For this reason, it is almost impossible to heat a house in
the disk because you can't stop the heat from escaping (p37 of the book The
Scientific Legacy of Poincare
2184718X/>).
I wonder then if it would similarly be almost impossible to surround
territory in hyperbolic Go. We need to try this!

Roice

--001a11408ca055e6d4052da12bd4
Content-Type: text/html; charset=UTF-8
Content-Transfer-Encoding: quoted-printable

--------------030604050500070100030608
Content-Type: text/plain; charset=utf-8; format=flowed
Content-Transfer-Encoding: quoted-printable

It was indeed exciting, and I'm even going to predict that it was a game=20
that will be remembered by history in much the same way as the pivotal=20
chess game in which IBM's chess bot Deep Blue beat the world champion=20
Gary Kasparov with a move so brilliant that he was convinced they had=20
cheated. This first of five Go games contained what appeared to me to be=20
a similarly devastating move by the machine. Here=20
is the video capture of the live=20
event. It's really ragged at the beginning as I don't think Google was=20
prepared for all the watchers but it gets better over time. Game=20
commentary is provided by a wonderful expert, Michael Redmond, who plays=20
at a similarly high level and he also explains a lot of basic concepts=20
though you can easily find many other great places to quickly learn the=20
basics if you are interested.

I've played one 13x13 game of Go on a torus and a another on a cylinder,=20
and they were very interesting. The problem with the torus, and perhaps=20
other polytopes, is that the lack of borders and corners leaves you=20
feeling rather naked as all territory must be built in empty space. It=20
was an equally strange experience going back to a normal board after=20
just one small game on a torus. I'm not sure how to describe the=20
experience but I'll just say that it hurt my normal game for a=20
surprising amount of time.

My game on the cylinder was a little more interesting to me. I think it=20
becomes natural for each player to sort of stake out one end, and then=20
to create rings in the middle in such a way as to capture opponent's=20
rings. You need to understand a bit about the game to understand this=20
but it seems to naturally come down to what are called capturing races.=20
I think that playing on a torus might work well if non-square dimensions=20
are chosen such that these sorts of rings become important but not too=20
important.

Go variants played on boards with different vertex valences have been=20
tried but I get the feeling that 4 really is the best choice. So Roice=20
may be right that a {5,4} would make for an interesting choice since it=20
preserves the familiar vertices. I don't think that infinite boards will=20
be attractive, but finite ones with negative curvature might work though=20
the lack of borders and corners might be even worse than on a torus.

MagicTile could be an ideal platform for testing out some of these=20
ideas. There is a small but passionate population of Go players who=20
enjoy exploring non-standard boards and would certainly love this idea.=20
I like the idea of playing on the skeleton of duoprisms though I think=20
the particular choice of duoprism will be very important to how well it=20
adapts to the game. If you implemented this, would it still work in the=20
3D view? As we've seen with the IRP puzzles, the 3D view was not helpful=20
but it sure looks great and helps to explain the topology. As with the=20
duoprisms, I suspect that the particular choice of IRP would be important.

If you're seriously considering integrating Go into MagicTile, I suggest=20
contacting some of the people in the Go community who like to play on=20
non-standard boards to find out what excites them the most.

-Melinda

On 3/9/2016 9:38 AM, Roice Nelson roice3@gmail.com [4D_Cubing] wrote:
>
>
> Anyone catch the match last night? Melinda and I did, and are=20
> enthusiastically discussing it. It was awesome! You can watch the=20
> remaining games live here=20
> .
>
> To connect the excitement back to the group, I wanted to mention you=20
> can play Go on the 1-skeletons of 4D polytopes using an early version=20
> of Jenn3D. Head to the very bottom of this page=20
> to try. Duoprisms are=20
> particularly interesting, because you can use them to make boards that=20
> remove all the edges of a traditional board but are otherwise the=20
> same. Playing on polytopes feels like it would generally have too=20
> much freedom though, especially if single stones have more than 4=20
> adjacent liberties.
>
> Adapting MagicTile to support Go might work well, since it would keep=20
> the boards as 2D surfaces. A {5,4} tiling would be a natural choice=20
> for a first board, and probably some of Andrea Hawksley's ideas about=20
> non-euclidean=20
> chess=20
> would apply. But I also wonder if hyperbolic Go would be=20
> fundamentally flawed. Random walks in the Poincar=C3=A9 disk inevitably=
=20
> escape to infinity. For this reason, it is almost impossible to heat=20
> a house in the disk because you can't stop the heat from escaping (p37=20
> of the book The Scientific Legacy of Poincare=20
> 082184718X/>).=20
> I wonder then if it would similarly be almost impossible to surround=20
> territory in hyperbolic Go. We need to try this!
>
> Roice


--------------030604050500070100030608
Content-Type: text/html; charset=utf-8
Content-Transfer-Encoding: quoted-printable



">


It was indeed exciting, and I'm even going to predict that it was a
game that will be remembered by history in much the same way as the
pivotal chess game in which IBM's chess bot Deep Blue beat the world
champion Gary Kasparov with a move so brilliant that he was
convinced they had cheated. This first of five Go games contained
what appeared to me to be a similarly devastating move by the
machine. Here is
the video capture of the live event. It's really ragged at the
beginning as I don't think Google was prepared for all the watchers
but it gets better over time. Game commentary is provided by a
wonderful expert, Michael Redmond, who plays at a similarly high
level and he also explains a lot of basic concepts though you can
easily find many other great places to quickly learn the basics if
you are interested.



I've played one 13x13 game of Go on a torus and a another on a
cylinder, and they were very interesting. The problem with the
torus, and perhaps other polytopes, is that the lack of borders and
corners leaves you feeling rather naked as all territory must be
built in empty space. It was an equally strange experience going
back to a normal board after just one small game on a torus. I'm not
sure how to describe the experience but I'll just say that it hurt
my normal game for a surprising amount of time.



My game on the cylinder was a little more interesting to me. I think
it becomes natural for each player to sort of stake out one end, and
then to create rings in the middle in such a way as to capture
opponent's rings. You need to understand a bit about the game to
understand this but it seems to naturally come down to what are
called capturing races. I think that playing on a torus might work
well if non-square dimensions are chosen such that these sorts of
rings become important but not too important.



Go variants played on boards with different vertex valences have
been tried but I get the feeling that 4 really is the best choice.
So Roice may be right that a {5,4} would make for an interesting
choice since it preserves the familiar vertices. I don't think that
infinite boards will be attractive, but finite ones with negative
curvature might work though the lack of borders and corners might be
even worse than on a torus.



MagicTile could be an ideal platform for testing out some of these
ideas. There is a small but passionate population of Go players who
enjoy exploring non-standard boards and would certainly love this
idea. I like the idea of playing on the skeleton of duoprisms though
I think the particular choice of duoprism will be very important to
how well it adapts to the game. If you implemented this, would it
still work in the 3D view? As we've seen with the IRP puzzles, the
3D view was not helpful but it sure looks great and helps to explain
the topology. As with the duoprisms, I suspect that the particular
choice of IRP would be important.



If you're seriously considering integrating Go into MagicTile, I
suggest contacting some of the people in the Go community who like
to play on non-standard boards to find out what excites them the
most.



-Melinda