Six Years Back, The Last-Ever Meat Victory vs A.I. In an Even Game

[Hal Womack 3-dan]

https://www.theatlantic.com/technology/archive/2016/03/the-invisible-opponent/475611/

How Google's AlphaGo Beat a Go World Champion
Inside a man-versus-machine showdown

By Christopher Moyer
MARCH 28, 2016

[The number of potential legal board positions is greater than the number of atoms in the universe.

...Demis Hassabis, DeepMind’s founder

...Other Korean professionals joke that they’re envious of Lee, that they feel the DeepMind Challenge Match is the easiest million dollars a top-level player could ever make.

... As you are reading this, AlphaGo is improving. It does not take breaks. It does not have days when it just doesn’t feel like practicing, days when it can’t kick its electronic brain into focus. Day in and day out, AlphaGo has been rocketing towards superiority, and the results are staggering.

...Gu Li, one of Lee’s long-term friends and rivals, comments on Chinese TV that Lee is fighting “a very lonely battle against an invisible opponent.”

...Then comes Lee’s move 78, which will come to be called his “Hand of God” move.

...In the end, finding no moves that improve its chances of winning, it [A.I.] begins playing nonsense moves, moves that actually reduce its own points. Finally, it resigns.

...Go is constantly evolving. What’s considered optimal play changes quickly. Humans have been honing our collective knowledge of the game for more than 2,500 years—the difference is that AlphaGo can do the same thing much, much faster.]
------------------------------------------------

I, Womack, call the A.I. program "TopBot" and the game itself "turff". Lee Sedol himself subsequently retired from playing professional _baduk_.

How much of a handicap would #1 meat need to beat TopBot? When, if ever, will we find out?

[sobriquet]

Post by Hal Womack 3-dan
https://www.theatlantic.com/technology/archive/2016/03/the-invisible-opponent/475611/
How Google's AlphaGo Beat a Go World Champion
Inside a man-versus-machine showdown
By Christopher Moyer
MARCH 28, 2016
[The number of potential legal board positions is greater than the number of atoms in the universe.
...Demis Hassabis, DeepMind’s founder
...Other Korean professionals joke that they’re envious of Lee, that they feel the DeepMind Challenge Match is the easiest million dollars a top-level player could ever make.
... As you are reading this, AlphaGo is improving. It does not take breaks. It does not have days when it just doesn’t feel like practicing, days when it can’t kick its electronic brain into focus. Day in and day out, AlphaGo has been rocketing towards superiority, and the results are staggering.
...Gu Li, one of Lee’s long-term friends and rivals, comments on Chinese TV that Lee is fighting “a very lonely battle against an invisible opponent.”
...Then comes Lee’s move 78, which will come to be called his “Hand of God” move.
...In the end, finding no moves that improve its chances of winning, it [A.I.] begins playing nonsense moves, moves that actually reduce its own points. Finally, it resigns.
...Go is constantly evolving. What’s considered optimal play changes quickly. Humans have been honing our collective knowledge of the game for more than 2,500 years—the difference is that AlphaGo can do the same thing much, much faster.]
------------------------------------------------
I, Womack, call the A.I. program "TopBot" and the game itself "turff". Lee Sedol himself subsequently retired from playing professional _baduk_.
How much of a handicap would #1 meat need to beat TopBot? When, if ever, will we find out?

Yet, bizarrely, we still don't have any AI yet that can explain go or that is even able to come up
with a basic concept like 'atari' or 'ladder' from scratch (that is, starting with just the rules of
go regarding the moves one is allowed to play).

Surely the art of teaching go is not that much more difficult than the art of playing go?
AI systems that can play go are much less useful than AI systems that can explain go
and boil it down to the essential concepts, to facilitate and optimize the learning process
to master go (both for AI systems learning from the experience of other AI systems, rather
than their own experience and for assisting humans to learn and understand go).

[Robert Jasiek]

Post by sobriquet
the art of teaching go is not that much more difficult than the art of playing go?

Of teaching very basic concepts, maybe. Advanced concepts require
sophisticated intelligence, for which AI is not ready.

[sobriquet]

Post by Robert Jasiek

Post by sobriquet
the art of teaching go is not that much more difficult than the art of playing go?

Of teaching very basic concepts, maybe. Advanced concepts require
sophisticated intelligence, for which AI is not ready.

Nonsense. AI has already demonstrated that it can easily beat people at the
game of go (regardless of their 'sophisticated' level of intelligence).
It's just that humans are rather clueless about the nature of intelligence and
that kind of explains why we don't have AI yet that can communicate at a
conceptual level.
Human intelligence is much more a matter of having huge numbers of people
clowning around and collectively accumulating scraps of knowledge over many
generations rather than single brilliant individual figuring out everything from
scratch (like the current AI approach, where a single AI system starts learning
from scratch all by itself to reach 9d level skills).

[Robert Jasiek]

Post by sobriquet
AI has already demonstrated that it can easily beat people at the
game of go

Of course, but this is not conceptual explanation in
human-understandable terms.

Post by sobriquet
It's just that humans are rather clueless about the nature of intelligence and
that kind of explains why we don't have AI yet that can communicate at a
conceptual level.

Not quite. Proof play, theorem application and expert systems are
examples of ways in which AI and humans can communicate in
human-understandable terms.

Post by sobriquet
Human intelligence is much more a matter of having huge numbers of people
clowning around and collectively accumulating scraps of knowledge over many
generations rather than single brilliant individual figuring out everything from
scratch (like the current AI approach

Human intelligence involves both: accumulating scraps of knowledge and
brilliant individuals' thinking. For the latter, read, e.g.,
Combinatorial Game Theory (by Siegel, quite a bit of the contents
applies to Go) and my book Endgame 5 - Mathematics, which contains 149
theorems and their proofs, of which 2/3 are mine.

[sobriquet]

Post by Robert Jasiek

Post by sobriquet
AI has already demonstrated that it can easily beat people at the
game of go

Of course, but this is not conceptual explanation in
human-understandable terms.

There is no reason whatsoever to assume that computers wouldn't excel at
that, if only humans figured out how to come up with a framework that allows
them to learn this skill.

Post by Robert Jasiek

Post by sobriquet
It's just that humans are rather clueless about the nature of intelligence and
that kind of explains why we don't have AI yet that can communicate at a
conceptual level.

Not quite. Proof play, theorem application and expert systems are
examples of ways in which AI and humans can communicate in
human-understandable terms.

It's certainly not hardware limitations that preclude computers from attaining
high-level conceptual information processing skills. So sooner or later it's bound
to arise (as computers transition from the skill or recognizing cats in pictures and
movies to reasoning about the concept of a cat).

Post by Robert Jasiek

Post by sobriquet
Human intelligence is much more a matter of having huge numbers of people
clowning around and collectively accumulating scraps of knowledge over many
generations rather than single brilliant individual figuring out everything from
scratch (like the current AI approach

Human intelligence involves both: accumulating scraps of knowledge and
brilliant individuals' thinking. For the latter, read, e.g.,
Combinatorial Game Theory (by Siegel, quite a bit of the contents
applies to Go) and my book Endgame 5 - Mathematics, which contains 149
theorems and their proofs, of which 2/3 are mine.

Books are outdated. I prefer videos. People's brilliancy can only come to fruition if they
build upon the enormous wealth of accumulated knowledge from their forebears.
You can't expect any individual human to come up with something like quantum field theory
by themselves, if they were raised by monkeys on a deserted island.
So it would be silly to expect such feats from AI systems gaining knowledge from
experience completely independently.

[Robert Jasiek]

Post by sobriquet
Books are outdated. I prefer videos.

Both have their purposes but describing books as outdated is improper.
Typically, books can have much denser contents than videos. If you
watch all go endgame videos, you can learn only 1% of all endgame
theory.

Post by sobriquet
People's brilliancy can only come to fruition if they
build upon the enormous wealth of accumulated knowledge from their forebears.

This is usually so but very much of my discoveries is built from
ground up.

Post by sobriquet
You can't expect any individual human to come up with something like quantum field theory
by themselves,

Einstein is an example of building the theory of relativity mostly by
himself seemingly out of nowhere. (It is possible though that his wife
contributed and Einstein might have failed to credit this; we cannot
know.)

IIRC, Boltzmann's entropy theory is another such example.

Quantum field theory just happens to be one of the countless common
examples of a theory developed by many.

Post by sobriquet
So it would be silly to expect such feats from AI systems gaining knowledge from
experience completely independently.

At least it would be silly to expect that AI would discover a
particular theory without prior directional guidance - instead it
might discover some entirely different theory in the future:)

[sobriquet]

Post by Robert Jasiek

Post by sobriquet
Books are outdated. I prefer videos.

Both have their purposes but describing books as outdated is improper.
Typically, books can have much denser contents than videos. If you
watch all go endgame videos, you can learn only 1% of all endgame
theory.

No way. Books are static. Videos are dynamic and can show animations.
That by itself makes books pretty much useless, compared to the power
of the video format (just like a picture can reveal more than a thousand
words, an animation can reveal more than a thousand pictures), but
interactive educational software can be an even way more powerful
format, as it can present information in an interactive way that
continually matches up the level of detail with background knowledge
of the user, to optimize the learning experience.

Post by Robert Jasiek

Post by sobriquet
People's brilliancy can only come to fruition if they
build upon the enormous wealth of accumulated knowledge from their forebears.

This is usually so but very much of my discoveries is built from
ground up.

Post by sobriquet
You can't expect any individual human to come up with something like quantum field theory
by themselves,

Einstein is an example of building the theory of relativity mostly by
himself seemingly out of nowhere. (It is possible though that his wife
contributed and Einstein might have failed to credit this; we cannot
know.)
IIRC, Boltzmann's entropy theory is another such example.
Quantum field theory just happens to be one of the countless common
examples of a theory developed by many.

That is the rule rather than the exception. People like Newton, Einstein and Boltzmann
are the exception and even they usually acknowledge they were very much
contributing some small but very significant pieces to the bulk of the puzzle
that was provided by others.

Post by Robert Jasiek

Post by sobriquet
So it would be silly to expect such feats from AI systems gaining knowledge from
experience completely independently.

At least it would be silly to expect that AI would discover a
particular theory without prior directional guidance - instead it
might discover some entirely different theory in the future:)

It would be nice to have a population of evolving AI systems with some selective
pressure towards the development of language, so they can learn both from their
own experience and the experience of their peers. Combining neural networks
and genetic algorithms somehow.

[Robert Jasiek]

Books are static. Videos are dynamic and can show animations.

[...]

interactive educational software can be an even way more powerful
format,

Usually, videos are also static in the sense that their contents is
defined on creation and does not change dynamically on viewing, quite
like printed books.

Books, videos and software can be used together one's flexible mind or
physical / virtual go boards.

The interaction of ebooks or software alone does not provide contents
that is not there. What matters the most is what and how much contents
a medium presents at all.

Dynamic go boards in ebooks or software allow studying variations and
their positions move after move that are not included or shown
explicitly. This may, or may not, be richer than printed books (e.g.,
some of my books show every relevant move or position and study and
explain ALL relevant variations and decisions.

For comparison, the most detailed ebooks I have seen study less, rely
on the reader to study all remaining relevant variations and decisions
on his own. Such ebooks show too few sample decisions so almost all
readers won't know how to find all missing relevant decisions on their
own. They can't be missing any relevant decisions in my related books
because they already contain ALL of them.

Furthermore, replaying moves and positions in ebooks or softwares
encourages lazy thinking instead of training visualisation, tactical
reading and positional judgement.

A technically more powerful because more dynamic format does not
guarantee better, or more complete, teaching.

[sobriquet]

Post by Robert Jasiek

Books are static. Videos are dynamic and can show animations.

[...]

interactive educational software can be an even way more powerful
format,

Usually, videos are also static in the sense that their contents is
defined on creation and does not change dynamically on viewing, quite
like printed books.
Books, videos and software can be used together one's flexible mind or
physical / virtual go boards.
The interaction of ebooks or software alone does not provide contents
that is not there. What matters the most is what and how much contents
a medium presents at all.
Dynamic go boards in ebooks or software allow studying variations and
their positions move after move that are not included or shown
explicitly. This may, or may not, be richer than printed books (e.g.,
some of my books show every relevant move or position and study and
explain ALL relevant variations and decisions.
For comparison, the most detailed ebooks I have seen study less, rely
on the reader to study all remaining relevant variations and decisions
on his own. Such ebooks show too few sample decisions so almost all
readers won't know how to find all missing relevant decisions on their
own. They can't be missing any relevant decisions in my related books
because they already contain ALL of them.
Furthermore, replaying moves and positions in ebooks or softwares
encourages lazy thinking instead of training visualisation, tactical
reading and positional judgement.
A technically more powerful because more dynamic format does not
guarantee better, or more complete, teaching.

You're right that books have been around much longer, so there is a
wealth of information available in books that has yet to become available
in other formats.
But AI developments go in leaps and bounds and we might not be that
far off from AI software that is able to crunch down a huge bunch of
books (in ebook format) to the essential conceptual structure and
makes this information accessible through a convenient interface that
can generate parts of this structure at an appropriate level of detail
based on the way the user has been interacting with it so far.
I don't think videos are somehow necessarily limited in scope and breadth
of the amount of detail they can cover, since you can make arbitrarily long
videos and split them up in convenient parts, though so far videos make
it hard to quickly locate very specific topics within a video. AI could
potentially help with that as well, so you can envision an AI system that
can quickly retrieve all the relevant sections of content in book or video
format when presented with a query to elucidate a particular aspect.

Ultimately, what matters most is that the person or system explaining the
topic has a very effective way of conveying that content in the most
insightful way, but video can be a more engaging and pleasing format
potentially (since personal presentation adds quite a bit of nuance, like
in the way people speak or their body language) as it is closer to natural
in-person interaction.
I'm also interested in learning content in general and topics like math and go are
good subjects that are quite abstract and challenging to master.

[Robert Jasiek]

Post by sobriquet
books have been around much longer, so there is a
wealth of information available in books that has yet to become available
in other formats.

It is not just the information that has been available for a long time
in books. In recent years, some go research has exploded in scope. In
the last 5 years, the amount of endgame evaluation theory has been
multiplied by a factor between 2 and 10.

Post by sobriquet
But AI developments go in leaps and bounds and we might not be that
far off from AI software that is able to crunch down a huge bunch of
books (in ebook format) to the essential conceptual structure and
makes this information accessible through a convenient interface that
can generate parts of this structure at an appropriate level of detail
based on the way the user has been interacting with it so far.

Possibly.

Post by sobriquet
I don't think videos are somehow necessarily limited in scope and breadth
of the amount of detail they can cover, since you can make arbitrarily long
videos and split them up in convenient parts,

As a wild first guess, presenting the contents of Endgame 5 -
Mathematics in videos presented well for the typical viewers would
result in at least 500 hours.

Post by sobriquet
Ultimately, what matters most is that the person or system explaining the
topic has a very effective way of conveying that content in the most
insightful way, but video can be a more engaging and pleasing format

I have viewed some videos explaining advanced physics. What such
videos do is to omit 99% of the theory and explain 1% to the
mathematically at least somewhat educated layman. This works because
the viewer does not need to learn more than he can from effortlessly
watching the videos. However, go skill is not like that. One cannot
only pick the cherries and expect to become much stronger. Instead,
one must also invest hard effort into skills and advanced theory.

Post by sobriquet
I'm also interested in learning content in general and topics like math and go are
good subjects that are quite abstract and challenging to master.

Precisely:)

[RichD]

Post by Robert Jasiek
In recent years, some go research has exploded in scope. In
the last 5 years, the amount of endgame evaluation theory has been
multiplied by a factor between 2 and 10.

Can you provide any such endgame examples, which would be
misplayed 5 years ago, where the new improved theory analyzes correctly?

--
Rich

[Robert Jasiek]

Post by RichD
Can you provide any such endgame examples, which would be
misplayed 5 years ago, where the new improved theory analyzes correctly?

The theory is not specifically designed for this purpose but rather is
designed to handle all ordinary examples correctly. Nevertheless,
there are some positions that fit your desire. Here is one from
Endgame 4 - Global Move Order, p. 52f, Example 9, refuting the
previous very popular advice of always seeking tedomari.

Finding this first counter-example took 4 days. Note that the largest
drop does not occur at the end, the starting player decides who gets
tedomari and his only correct choice gives tedomari to the opponent.
All drops play their role and we must make the right decision at
branches, which is the correct theory here. Studying variations,
resulting counts and related decisions at branches is an exercise.

X X X O O O O . O . O X X X . X
X X X O X X O O O O O O O X X X X
X X X O X X O X X O X X O O O O X
X X X O X X O X X O X O O O X . X
X . O O X X O X X O X O X O O O X
. O O O X X O X X O X O X O X O X
X X O O X X O X X O X O X O X O X
X X X O X X O X X O X O X O X O X
X X X O X X O X X O X O X O X O X
X X X O X X O X X O X O X O X O X
. O O O . O O . O O . O . O . O X
X X X X X X X X X X X X X X X X X

Move value M = 18.25.
Follow-up move value F = 13.5.
Environment's move values 18, 16, 9, 6, 5, 1.
Correct order: 18.25 - 18 - 16 - 13.5 (opponent's tedomari) - 9 -...
Resulting count = -77.

[Robert Jasiek]

Let me continue endgame examples that were, or often were, misjudged 5
years ago but can be judged correctly with the new theory.

A frequent cause of misjudgement is wrongly assessing the type of a
local endgame. Previously, almost nobody knew how to verify the type
but everybody just guessed it, only Bill Spight seemed to know and
thermography circumvented types by slopes of mappings. However,
graphical or algebraic thermography is a tool for mathematicians
developed and refined during the previous decades and essentially
inapplicable while playing a game. My book Endgame 3 - Accurate Local
Evaluation has made Bill's ideas for identifying and verifying types
available for everybody. This is Example 1 on p. 6.

X X X X O .
. O O O O .
X X . O . .
. O O O O O
X X X X X O
. . . . X .

Determine the move value! (Use modern endgame theory aka miai counting
aka move value per excess play.) To determine the correct move value
for the purpose of avoiding mistakes in your games, determine the
correct type local gote or local sente or "ambiguous". Previously,
amateurs and professionals have made frequent mistakes for such
questions applied to the most ordinary shapes.

[Robert Jasiek]

Hint:

X X X X O .
. O O O O .
X X . O . .
X O O O O O
X X X X X O
. . . . X .

Black follower's count B = -4

(This is the average of 0 and -8.)

X X X X O .
. O O O O .
X X . O . .
O O O O O O
X X X X X O
. . . . X .

White follower's count W = -14

[Robert Jasiek]

Hint 2:

X X X X O .
. O O O O .
X X . O . .
X O O O O O
X X X X X O
. . . . X .

The follow-up move value is F = (0 - (-8)) / 2 = 4.

[Robert Jasiek]

Is everybody still struggling to decide whether it is a local gote or
sente?

X X X X O .
. O O O O .
X X . O . .
. O O O O O
X X X X X O
. . . . X .

initial position

X X X X O .
. O O O O .
X X . O . .
X O O O O O
X X X X X O
. . . . X .

Black follower's count B = -4

X X X X O .
. O O O O .
X X . O . .
O O O O O O
X X X X X O
. . . . X .

White follower's count W = -14


Hints 3:


. . . . O .
O O O O O .
X X . O . .
X O O O O O
X X X X X O
. . . . X .

Territory difference = -4. Prisoner difference = -4.
Sente follower's count S = -8.


X X X X O .
. O O O O .
X X . O . .
. O O O O O
X X X X X O
. . . . X .

initial position (repeated)


Initial position's and Black's tentative sente move value
M_sente = S - W = -8 - (-14) = 6.


Initial position's tentative gote move value
M_gote = (B - W) / 2 = (-4 - (-14)) / 2 = 5.


Which is the move value? Why? For the reason, you can choose from four
alternative, correct answers.

[Robert Jasiek]

X X X X O .
. O O O O .
X X . O . .
. O O O O O
X X X X X O
. . . . X .

We have calculated

Black's tentative sente move value M_sente = 6,

tentative gote move value M_gote = 5,

(Black's) follow-up move value F = 4.


(White's follow-up move value, that is the follow-up move value after
White's first play, is 0.)


The initial local endgame belongs to the type of local endgames in
which a gote sequence might be continued as a sente sequence. The
initial local endgame is a local gote with the move value

M := M_gote = 5 because


M_sente > M_gote > F <=> 6 > 5 > 4.


The three conditions are equivalent so we can verify the type local
gote by each of them:

M_sente > M_gote

<=>

M_sente > F

<=>

M_gote > F.

Black's tentative sente move value is larger than the tentative gote
move value.

Black's tentative sente move value is larger than the follow-up move
value.

The tentative gote move value is larger than the follow-up move value.


A local gote is characterised by decreasing move values: its move
value M is larger than its follow-up move value F.


We must not trust a visual impression of threatening to connect the
larger string. Such does not necessarily mean a local sente. Instead,
we determine the type of the local endgame by a value condition.

It is insufficient to calculate some tentative move value. We must
also verify or refute it by a value condition. Almost all teachers and
players do it wrongly by only calculating some tentative move value
and guessing that it would have the correct type and be the correct
move value. Except for earlier thermography and Bill Spight's earlier
insight, correct evaluation including verification has only arisen
during the last 5 years.


X X X X O . O O
. O O O O . O O
X X . O . . O .
. O O O O O O X
X X X X X O O X
. . . . X . X X

Black to move.

The simple gote on the right side has the gote move value 4.5.

X X X X O . O O
. O O O O . O O
X X . O . . O .
1 O O O O O O X
X X X X X O O X
. . . . X 2 X X

As we have determined, the left local endgame is a local gote. It can
be correct to reply elsewhere because the follow-up move value can be
smaller than a move value elsewhere on the board, as here.