From: Ben Goertzel
Sent: Monday, June 28, 2010 12:03 AM
To: agi
Subject: Re: [agi] Huge Progress on the Core of AGI
Even with the variations you mention, I remain highly
confident this is not a difficult problem for narrow-AI
machine learning methods
-- Ben G
On Sun, Jun 27, 2010 at 6:24 PM, Mike Tintner
<[email protected]> wrote:
I think you're thinking of a plodding limited-movement
classic Pong line.
I'm thinking of a line that can like a human
player move with varying speed and pauses to more or
less any part of its court to hit the ball, and then
hit it with varying speed to more or less any part of
the opposite court. I think you'll find that bumps up
the variables if not unknowns massively.
Plus just about every shot exchange presents you with
dilemmas of how to place your shot and then move in
anticipation of your opponent's return .
Remember the object here is to present a would-be AGI
with a simple but *unpredictable* object to deal with,
reflecting the realities of there being a great many
such objects in the real world - as distinct from
Dave's all too predictable objects.
The possible weakness of this pong example is that
there might at some point cease to be unknowns, as
there always are in real world situations, incl
tennis. One could always introduce them if necessary -
allowing say creative spins on the ball.
But I doubt that it will be necessary here for the
purposes of anyone like Dave - and v. offhand and
with no doubt extreme license this strikes me as not a
million miles from a hyper version of the TSP problem,
where the towns can move around, and you can't be sure
whether they'll be there when you arrive. Or is there
an "obviously true" solution for that problem too?
[Very convenient these obviously true solutions].
From: Jim Bromer
Sent: Sunday, June 27, 2010 8:53 PM
To: agi
Subject: Re: [agi] Huge Progress on the Core of AGI
Ben: I'm quite sure a simple narrow AI system could
be constructed to beat humans at Pong ;p
Mike: Well, Ben, I'm glad you're "quite sure" because
you haven't given a single reason why.
Although Ben would have to give us an actual example
(of a pong program that could beat humans at
Pong) just to make sure that it is not that difficult
a task, it seems like such an obviously true statement
that there is almost no incentive for anyone to try
it. However, there are chess programs that can beat
the majority of people who play chess without outside
assistance.
Jim Bromer
On Sun, Jun 27, 2010 at 3:43 PM, Mike Tintner
<[email protected]> wrote:
Well, Ben, I'm glad you're "quite sure"
because you haven't given a single reason why.
Clearly you should be Number One advisor on
every Olympic team, because you've cracked the
AGI problem of how to deal with opponents that
can move (whether themselves or balls) in
multiple, unpredictable directions, that is at
the centre of just about every field and court
sport.
I think if you actually analyse it, you'll
find that you can't predict and prepare for
the presumably at least 50 to 100 spots on a
table tennis board/ tennis court that your
opponent can hit the ball to, let
alone for how he will play subsequent 10 to 20
shot rallies - and you can't devise a
deterministic program to play here. These are
true, multiple-/poly-solution problems rather
than the single solution ones you are familiar
with.
That's why all of these sports have normally
hundreds of different competing philosophies
and strategies, - and people continually can
and do come up with new approaches and styles
of play to the sports overall - there are
endless possibilities.
I suspect you may not play these sports,
because one factor you've obviously ignored
(although I stressed it) is not just the
complexity but that in sports players can and
do change their strategies - and that would
have to be a given in our computer game. In
real world activities, you're normally
*supposed* to act unpredictably at least some
of the time. It's a fundamental subgoal.
In sport, as in investment, "past performance
is not a [sure] guide to future performance" -
companies and markets may not continue to
behave as they did in the past - so that
alone buggers any narrow AI predictive
approach.
P.S. But the most basic reality of these
sports is that you can't cover every shot or
move your opponent may make, and that gives
rise to a continuing stream of genuine
dilemmas . For example, you have just returned
a ball from the extreme, far left of your
court - do you now start moving rapidly
towards the centre of the court so that you
will be prepared to cover a ball to the
extreme, near right side - or do you move more
slowly? If you don't move rapidly, you won't
be able to cover that ball if it comes. But if
you do move rapidly, your opponent can play
the ball back to the extreme left and catch
you out.
It's a genuine dilemma and gamble - just like
deciding whether to invest in shares. And
competitive sports are built on such
dilemmas.
Welcome to the real world of AGI problems. You
should get to know it.
And as this example (and my rock wall
problem) indicate, these problems can
be as simple and accessible as fairly easy
narrow AI problems.
From: Ben Goertzel
Sent: Sunday, June 27, 2010 7:33 PM
To: agi
Subject: Re: [agi] Huge Progress on the Core
of AGI
That's a rather bizarre suggestion Mike ...
I'm quite sure a simple narrow AI system could
be constructed to beat humans at Pong ;p ...
without teaching us much of anything about
intelligence...
Very likely a narrow-AI machine learning
system could *learn* by experience to beat
humans at Pong ... also without teaching us
much
of anything about intelligence...
Pong is almost surely a "toy domain" ...
ben g
On Sun, Jun 27, 2010 at 2:12 PM, Mike Tintner
<[email protected]> wrote:
Try ping-pong - as per the computer
game. Just a line (/bat) and a
square(/ball) representing your
opponent - and you have a line(/bat)
to play against them
Now you've got a relatively simple
true AGI visual problem - because if
the opponent returns the ball somewhat
as a real human AGI does, (without
the complexities of spin etc just
presumably repeatedly changing the
direction (and perhaps the speed) of
the returned ball) - then you have a
fundamentally *unpredictable* object.
How will your program learn to play
that opponent - bearing in mind that
the opponent is likely to keep
changing and even evolving
strategy? Your approach will have to
be fundamentally different from how a
program learns to play a board game,
where all the possibilities are
predictable. In the real world, "past
performance is not a [sure] guide to
future performance". Bayes doesn't
apply.
That's the real issue here - it's not
one of simplicity/complexity - it's
that your chosen worlds all consist
of objects that are predictable,
because they behave consistently, are
shaped consistently, and come in
consistent, closed sets - and can
only basically behave in one way at
any given point. AGI is about dealing
with the real world of objects that
are unpredictable because they behave
inconsistently,even contradictorily,
are shaped inconsistently and come in
inconsistent, open sets - and can
behave in multi-/poly-ways at any
given point. These differences apply
at all levels from the most complex to
the simplest.
Dealing with consistent (and regular)
objects is no preparation for dealing
with inconsistent, irregular
objects.It's a fundamental error
Real AGI animals and humans were
clearly designed to deal with a world
of objects that have some
consistencies but overall are
inconsistent, irregular and come in
open sets. The perfect regularities
and consistencies of geometrical
figures and mechanical motion (and
boxes moving across a screen) were
only invented very recently.
From: David Jones
Sent: Sunday, June 27, 2010 5:57 PM
To: agi
Subject: Re: [agi] Huge Progress on
the Core of AGI
Jim,
Two things.
1) If the method I have suggested
works for the most simple case, it is
quite straight forward to add
complexity and then ask, how do I
solve it now. If you can't solve that
case, there is no way in hell you will
solve the full AGI problem. This is
how I intend to figure out how to
solve such a massive problem. You
cannot tackle the whole thing all at
once. I've tried it and it doesn't
work because you can't focus on
anything. It is like a Rubik's cube.
You turn one piece to get the color
orange in place, but at the same time
you are screwing up the other colors.
Now imagine that times 1000. You
simply can't do it. So, you start with
a simple demonstration of the
difficulties and show how to solve a
small puzzle, such as a Rubik's cube
with 4 little cubes to a side instead
of 6. Then you can show how to solve 2
sides of a rubiks cube, etc.
Eventually, it will be clear how to
solve the whole problem because by the
time you're done, you have a complete
understanding of what is going on and
how to go about solving it.
2) I haven't mentioned a method for
matching expected behavior to
observations and bypassing the default
algorithms, but I have figured out
quite a lot about how to do it. I'll
give you an example from my own notes
below. What I've realized is that the
AI creates *expectations* (again).
When those expectations are matched,
the AI does not do its default
processing and analysis. It doesn't do
the default matching that it normally
does when it has no other knowledge.
It starts with an existing hypothesis.
When unexpected observations or
inconsistencies occur, then the AI
will have a *reason* or *cue* (these
words again... very important
concepts) to look for a better
hypothesis. Only then, should it look
for another hypothesis.
My notes:
How does the ai learn and figure out
how to explain complex unforseen
behaviors that are not
preprogrammable. For example the
situation above regarding two windows.
How does it learn the following
knowledge: the notepad icon opens a
new notepad window and that two
windows can exist... not just one
window that changes. the bar with the
notepad icon represenants an instance.
the bar at the bottom with numbers on
it represents multiple instances of
the same window and if you click on it
it shows you representative bars for
each window.
How do we add and combine this
complex behavior learning,
explanation, recognition and
understanding into our system?
Answer: The way that such things are
learned is by making observations,
learning patterns and then connecting
the patterns in a way that is
consistent, explanatory and likely.
Example: Clicking the notepad icon
causes a notepad window to appear with
no content. If we previously had a
notepad window open, it may seem like
clicking the icon just clears the
content by the instance is the same.
But, this cannot be the case because
if we click the icon when no notepad
window previously existed, it will be
blank. based on these two experiences
we can construct an explanatory
hypothesis such that: clicking the
icon simply opens a blank window. We
also get evidence for this conclusion
when we see the two windows side by
side. If we see the old window with
the content still intact we will
realize that clicking the icon did not
seem to have cleared it.
Dave
On Sun, Jun 27, 2010 at 12:39 PM, Jim
Bromer <[email protected]> wrote:
On Sun, Jun 27, 2010 at 11:56
AM, Mike Tintner
<[email protected]>
wrote:
Jim :This illustrates
one of the things
wrong with the
dreary instantiations
of the prevailing mind
set of a group. It is
only a matter of time
until you discover
(through experiment)
how absurd it is to
celebrate the triumph
of an overly
simplistic solution to
a problem that is, by
its very potential,
full of possibilities]
To put it more
succinctly, Dave & Ben
& Hutter are doing the
wrong subject - narrow
AI. Looking for the
one right prediction/
explanation is narrow
AI. Being able to
generate more and more
possible explanations,
wh. could all be
valid, is AGI. The
former is rational,
uniform thinking. The
latter is creative,
polyform thinking. Or,
if you prefer, it's
convergent vs
divergent thinking,
the difference between
wh. still seems to
escape Dave & Ben &
most AGI-ers.
Well, I agree with what (I
think) Mike was trying to get
at, except that I understood
that Ben, Hutter and
especially David were not only
talking about prediction as a
specification of a single
prediction when many possible
predictions (ie expectations)
were appropriate for
consideration.
For some reason none of you
seem to ever talk about
methods that could be used to
react to a situation with the
flexibility to integrate the
recognition of different
combinations of familiar
events and to classify unusual
events so they could be
interpreted as more familiar
*kinds* of events or as novel
forms of events which might be
then be integrated. For me,
that seems to be one of the
unsolved problems. Being able
to say that the squares move
to the right in unison is a
better description than saying
the squares are dancing the
irish jig is not really
cutting edge.
As far as David's comment that
he was only dealing with the
"core issues," I am sorry but
you were not dealing with the
core issues of contemporary
AGI programming. You were
dealing with a primitive
problem that has been
considered for many years, but
it is not a core research
issue. Yes we have to work
with simple examples to
explain what we are talking
about, but there is a
difference between an abstract
problem that may be central to
your recent work and a core
research issue that hasn't
really been solved.
The entire problem of dealing
with complicated situations is
that these narrow AI methods
haven't really worked. That
is the core issue.
Jim Bromer
agi | Archives
| Modify Your
Subscription
agi | Archives |
Modify Your
Subscription
agi | Archives |
Modify Your
Subscription
--
Ben Goertzel, PhD
CEO, Novamente LLC and Biomind LLC
CTO, Genescient Corp
Vice Chairman, Humanity+
Advisor, Singularity University and
Singularity Institute
External Research Professor, Xiamen
University, China
[email protected]
"
“When nothing seems to help, I go look at a
stonecutter hammering away at his rock,
perhaps a hundred times without as much as a
crack showing in it. Yet at the hundred and
first blow it will split in two, and I know it
was not that blow that did it, but all that
had gone before.”
agi | Archives |
Modify Your
Subscription
agi | Archives |
Modify Your
Subscription
agi | Archives | Modify
Your Subscription
agi | Archives | Modify
Your Subscription
--
Ben Goertzel, PhD
CEO, Novamente LLC and Biomind LLC
CTO, Genescient Corp
Vice Chairman, Humanity+
Advisor, Singularity University and Singularity Institute
External Research Professor, Xiamen University, China
[email protected]
"
“When nothing seems to help, I go look at a stonecutter
hammering away at his rock, perhaps a hundred times without as
much as a crack showing in it. Yet at the hundred and first
blow it will split in two, and I know it was not that blow
that did it, but all that had gone before.”
agi | Archives | Modify Your
Subscription
agi | Archives | Modify Your
Subscription
--
Ben Goertzel, PhD
CEO, Novamente LLC and Biomind LLC
CTO, Genescient Corp
Vice Chairman, Humanity+
Advisor, Singularity University and Singularity Institute
External Research Professor, Xiamen University, China
[email protected]
"
“When nothing seems to help, I go look at a stonecutter hammering away
at his rock, perhaps a hundred times without as much as a crack
showing in it. Yet at the hundred and first blow it will split in two,
and I know it was not that blow that did it, but all that had gone
before.”
agi | Archives | Modify Your
Subscription
