[agi] Re: Solomonoff Machines – up close and personal

[Shane Legg]

Hi Ed,

So is the real significance of the universal prior, not its probability
> value given in a given probability space (which seems relatively
> unimportant, provided is not one or close to zero), but rather the fact that
> it can model almost any kind of probability space?
>

It just takes a binary string as input.  If you can express your problem as
one in which
a binary string represents what has been observed so far, and a continuation
of this
string represents what happens next, then Solomonoff induction can deal with
it.
So you don't have to "pick the space".  You do however have to take your
problem
and represent it as binary data and feed it in, just as you do when you put
any kind
of data into a computer.

The power of the universal prior comes from the fact that it takes all
computable
distributions into account.  In a sense it contains all well defined
hypotheses about
what the structure in the string could be.  This is a point that is worth
contemplating
for awhile.  If there is any structure in there and this structure can be
described by
a program on a computer, even a probabilistic one, then it's already
factored into
the universal prior and the Solomonoff predictor is already taking it into
account.

How does the Kolmogorov complexity help deal with this problem?
>

The key thing that Kolmogorov complexity provides is that it assigns a
weighting to
each hypothesis in the universal prior that is inversely proportional to the
complexity
of the hypothesis.  This means that the Solomonoff predictor respects, in
some sense,
the principle of Occam's razor.  That is, a priori, simpler things are
considered more
likely than complex ones.

ED##> ??Shane??, what are the major ways programs are used in a
> Solomonoff machine?  Are they used for generating and matching patterns? Are
> they used for generating and creating context specific instantiations of
> behavioral patterns?
>
Keep in mind that Solomonoff induction is not computable.  It is not an
algorithm.
The role that programs play is that they are used to "construct" the
universal prior.
Once this is done, the Solomonoff predictor just takes the prior and
conditions on
the observed string so far to work out the distribution over the next bit.
That's all.


>Lukasz##> The programs are generally required to exactly match in AIXI
> (but not in AIXItl I think).
> ED##> ??Shane??, could you please give us an assist on this one? Is
> exact matching required?  And if so, is this something that could be
> loosened in a real machine?
>
Exact pattern matching is required in the sense that if a hypothesis says
that
something cannot happen, and it does, then that hypothesis is effectively
discarded.

A real machine might have to loosen this, and many other things.  Note that
nobody
I know is trying to build a real AGI machine based on Solomonoff's model.


Isn't there a large similarity between a Solomonoff machine that could learn
> a hierarchy of pattern representing programs and Jeff Hawking's hierarchical
> learning (as represented in the Serre paper).  One could consider the
> patterns at each level of the higherarchy as sub-routines.  The system is
> designed to increase its representational efficiency by having
> representational subroutines available for use by multiple different
> patterns at higher compositional levels.  To the extent that a MOSES-type
> evolutionary system could be set to work making such representations more
> compact, it would become clear how semi-Solomonoff machines could be made to
> work in the practical world.
>

In think the point is that if you can do really really good general sequence
prediction (via something
impractical like Solomonoff induction, or practical like the cortex) then
you're a long way towards
being able to build a pretty impressive AGI.  Some of Hutter's students are
interested in the latter.



> The def of Solomonoff induction on the web and even in Shane Legg's paper
> "Solomonoff induction" make it sound like it is merely Bayesian induction,
> using the picking of priors based on Kolmogorov complexity.
>
Yes, that's all it is.

But statements made by Shane and Lukasz appears to imply that a Solomonoff
> machine uses programming and programming size as a tool for pattern
> representation, generalization, learning, inference, and more.
>
All these programs are weighted into that universal prior.


> So I think (but I could well be wrong) I know what that means.
> Unfortunately I am a little fuzzy about whether NCD would take "what"
> information, "what-with-what" or binding information, or frequency
> information sufficiently into account to be an optimal measure of
> similarity.  Is this correct?
>
NCD is just a computable approximation.  The universal similarity metric (in
the Li and Vitanyi book
that I cited) gives the pure incomputable version.  The pure version
basically takes all effective
similarity metrics into account when working out how similar two things
are.  So if you h

[agi] Re: How valuable is Solmononoff Induction for real world AGI?

[Shane Legg]

Hello Edward,

I'm glad you found some of the writing and links interesting.  Let me try to
answer some of your questions.

> I understand the basic idea that if you are seeking a prior likelihood for
> the occurrence of an event and you have no data about the frequency of its
> occurrence -- absent any other knowledge -- some notion of the complexity,
> in information-theory terms, of the event might help you make a better
> guess.  This makes sense because reality is a big computer, and complexity
> -- in terms of the number of combined events required to make reality cough
> up a given event , and the complexity of the space in which those events are
> to be combined -- should to some extent be related to the event's
> probability.  I can understand how such complexity could be approximated by
> the length of code required in some a theoretical Universal computer to
> model such real world event-occurrence complexity.
>
This seems like a reasonable summary to me.

Ok, let's take your example as I think it captures the essence of what
you are getting at:

> So what I am saying is, for example, that if you are receiving a sequence
> of bytes from a video camera, much of the complexity in the input stream
> might not be related to complexity-of-event-creation- or Occam's-razor-type
> issues, but rather to complexity of perception, or similarity understanding,
> or of appropriate context selection, factors which are not themselves
> necessarily related to complexity of occurrence.
>
In short, yes.  So, for example, if you hooked up a Solomonoff induction
machine to
a video camera it would first need to, in some sense, "understand" the
nature of this
input stream.  This may be more complex than what the camera is actually
looking at!

Given that the Solomonoff machine starts from zero knowledge of the world,
other than
a special form of prior knowledge provided by the universal prior, there is
no way around
this problem.  Somehow it has to learn this stuff.  The good news, as
Solomonoff proved,
is that if the encoding of the video input stream isn't too crazy complex (
i.e. there exists
a good algorithm to process the stream that isn't too long), then the
Solomonoff machine
will very quickly work out how to understand the video stream.  Furthermore,
if we put
such a camera on a robot or something wandering around in the world, then it
would not
take long at all before the complexity of the observed world far surpassed
the complexity
of the video stream encoding.

Perhaps what you should consider is a Solomonoff machine that has been
pre-trained
to do, say, vision.  That is, you get the machine and train it up on some
simple vision
input so that it understands the nature of this input.  Only then do you
look at how well
it performs at finding structure in the world though its visual input
stream.

Furthermore, I am saying that for an AGI it seems to me it would make much
> more sense to attempt to derive priors from notions of similarity, of
> probabilities of similar things, events, and contexts, and from things like
> causal models for similar or generalized classes.  There is usually much
> from reality that we do know that we can, and do, use when learning about
> things we don't know.
>
Yes.  In essence what you seem to be saying is that our prior knowledge of
the
world strongly biases how we interpret new information.  So, to use your
example,
we all know that people living on a small isolated island are probably
genetically quite
similar to each other.  Thus, if we see that one has brown skin, we will
guess that the
others probably do also.  However, weight is not so closely tied to
genetics, and so if
one is obese then this does not tell us much about how much other islanders
weigh.

Out-of-the-box a Solomonoff machine doesn't know anything about genetics and
weight, so it can't make such inferences based on seeing just one islander.
However,
if it did have prior experience with genetics etc., then it too would
generalise as you
describe using context.  Perhaps the best place to understand the theory of
this
is section 8.3 from "An Introduction to Kolmogorov complexity and its
Applications"
by Li and Vitanyi.  You can also find some approximations to this theory
that have
been applied in practice to many diverse problems under the title of
"Normalized
Compression Distance" or NCD.  A lot of this work has been done by Rudi
Cilibrasi.

 HOW VALUABLE IS SOLMONONOFF INDUCTION FOR REAL WORLD AGI?
>
Well ;-)

In a certain literal sense, not much, as it is not computable.  However,
many
practical methods in machine learning and statistics can be viewed as
computable
approximations of Solomonoff induction, and things like NCD have been used
in practice with some success.  And who knows, perhaps some very smart
person will come up with a new version of Solomonoff induction that is much
more practically useful.  Personally, I suspect other approaches will reach
human
level AGI first.

If you are interested in t

Re: [agi] How can you prove form/behaviour are disordered?

[Shane Legg]

On 6/8/07, Matt Mahoney <[EMAIL PROTECTED]> wrote:


"The author has received reliable information, from a Source who wishes to
> remain anonymous, that the decimal expansion of Omega begins
>
> Omega = 0.998020554253273471801908..."

For which choice of universal Turing machine?




It's actually  0.00787499699781238...

At least when based on the Turing machine described here:

http://www.emis.de/journals/EM/expmath/volumes/11/11.3/Calude361_370.pdf

Shane

-
This list is sponsored by AGIRI: http://www.agiri.org/email
To unsubscribe or change your options, please go to:
http://v2.listbox.com/member/?member_id=231415&user_secret=e9e40a7e

Re: [agi] NARS: definition of intelligence

[Shane Legg]

Pei,

Yes, the book is the best source for most of the topics. Sorry for the

absurd price, which I have no way to influence.



It's $190.  Somebody is making a lot of money on each copy and
I'm sure it's not you.  To get a 400 page hard cover published at
lulu.com is more like $25.

Shane

-
This list is sponsored by AGIRI: http://www.agiri.org/email
To unsubscribe or change your options, please go to:
http://v2.listbox.com/member/?member_id=231415&user_secret=e9e40a7e

Re: [agi] Intelligence vs Efficient Intelligence

[Shane Legg]

Matt,

Shane Legg's definition of universal intelligence requires (I believe)

complexity but not adaptability.


In a universal intelligence test the agent never knows what the environment
it is facing is.  It can only try to learn from experience and adapt in
order to
perform well.  This means that a system which is not adaptive will have a
very low universal intelligence.  Even within a single environment, some
environments will change over time and thus the agent must adapt in order
to keep performing well.

Shane

-
This list is sponsored by AGIRI: http://www.agiri.org/email
To unsubscribe or change your options, please go to:
http://v2.listbox.com/member/?member_id=231415&user_secret=fabd7936

Re: [agi] definitions of intelligence, again?!

[Shane Legg]

On 5/17/07, Pei Wang <[EMAIL PROTECTED]> wrote:


Sorry, it should be "I assume you are not arguing that evolution is
the only way to produce intelligence ...".



Definitely not.  Though the results in my elegant sequence prediction
paper show that at some point math is of no further use due to Goedel
incompleteness.  To go beyond that point things like evolution may be
necessary.

Shane

-
This list is sponsored by AGIRI: http://www.agiri.org/email
To unsubscribe or change your options, please go to:
http://v2.listbox.com/member/?member_id=231415&user_secret=fabd7936

Re: [agi] definitions of intelligence, again?!

[Shane Legg]

Mike,

And how do you measure how they deal with fear and uncertainty? How much

time do you allow the agent for hesitation, changing their mind, chopping
and changing their mind, and staring blankly into space? How do you compute
their "dark nights of the soul" - as Pei and every creative person will
have? These are parts of performance too.



I simply don't care whether an algorithm experiences "dark nights of the
soul",
whether it is afraid, uncertain, changes its mind, is greedy etc.  All I
care about
is end performance.  The same is true of the soul, consciousness, true love
etc.

Take consciousness.  Does consciousness improve performance in some
situations?  If it does then universal intelligence tests for consciousness
because it will pick up this performance  advantage in these situations.

Or, if consciousness has no impact on a system's performance in any
situation, then from my perspective whether a system is conscious or
not has nothing to do with its intelligence.  It may be the case that some
very intelligent systems are not conscious.  I don't know how to measure
consciousness, so I can't say for sure.

Shane

-
This list is sponsored by AGIRI: http://www.agiri.org/email
To unsubscribe or change your options, please go to:
http://v2.listbox.com/member/?member_id=231415&user_secret=fabd7936

Re: [agi] definitions of intelligence, again?!

[Shane Legg]

Mike,

There is no right way to deal with problematic problems, period (no right
way to write that essay or work out that marketing strategy). There are
reasonably effective and unreasonably uneffective ways, but no right ways,
and you can't be sure in advance which will and which won't be more or less
effective.

In defining universal intelligence I have not assumed that there is a right
way to
write an essay, or work out a marketing strategy, or anything else.  All I
ask is,
if you define a problem and measure performance according to something, how
well can the system perform with respect to that measure?  How the agent
goes
about doing this is left open.

Shane

-
This list is sponsored by AGIRI: http://www.agiri.org/email
To unsubscribe or change your options, please go to:
http://v2.listbox.com/member/?member_id=231415&user_secret=fabd7936

Re: [agi] Intelligence vs Efficient Intelligence

[Shane Legg]

Ben,

According to this distinction, AIXI and evolution have high intelligence

but low efficient intelligence.



Yes, and in the case of AIXI it is presumably zero given that the resource
consumption is infinite.  Evolution on the other hand is just efficient
enough
that when implemented on a crazy enough scale the results can be pretty
amazing.


If this hypothesis is correct then AIXI and the like don't really tell us

much about
what matters, which is the achievement of efficient intelligence in
relevant real-world
contexts...



That might well be true.

I don't want to give the impression that I don't care about the efficiency
of
intelligence.  On any given hardware the most intelligent system will be the
one that runs the algorithm with the greatest intelligence efficiency.
Thus,
if I want to see very intelligent systems then I need to care about how
efficient they are.  Nevertheless, it is still the end product raw
intelligence
generated by the system that really excites me, rather than statistics on
its
internal efficiency.

Shane

-
This list is sponsored by AGIRI: http://www.agiri.org/email
To unsubscribe or change your options, please go to:
http://v2.listbox.com/member/?member_id=231415&user_secret=fabd7936

Re: [agi] definitions of intelligence, again?!

[Shane Legg]

Pei,

This just shows the complexity of "the usual meaning of the word

intelligence" --- many people do associate with the ability of solving
hard problems, but at the same time, many people (often the same
people!) don't think a brute-force solution show any intelligence.



I think this comes from the idea people have that things like intelligence
and creativity must derive from some very clever process.  A relatively
dumb process implemented on a mind blowingly vast scale intuitively
doesn't seem like it could be sufficient.

I think the intelligent design movement gets its strength from this
intuition.
People think, "How could something as complex and amazing as the human
body and brain come out of not much more than random coin flips?!?!?"
They figure that the algorithm of evolution is just too simple and therefore
dumb to do something as amazing as coming up with the human brain.
Only something with super human intelligence could achieve such a thing.

The solution I'm proposing is that we consider that relatively simple rules
when implemented on sufficiently vast scales can be very intelligent.  From
this perspective, humans are indeed the product of intelligence, but the
intelligence isn't God's, its a 4 billion year global scale evolutionary
process.


When "intelligence" is used on human, there is no problem, since few

hard problem can be solved by the human mind by brute-force.



Maybe humans are a kind of brute-force algorithm?  Perhaps the
important information processing that takes place in neurons etc.
is not all that complex, the amazing power of the system largely
comes from its gigantic scale?




At this point, you see "capability" as more essential, while
I see "adaptivity" as more essential.



Yes, I take capability as primary.  However, adaptivity is implied
by the fact that being adaptable makes a system more capable.


today, conventional computers

solve many problems better than the human mind, but I don't take that
as reason for them to be more intelligent.



The reason for that, I believe, is because the set of problems that they
can solve is far too narrow.  If they were able to solve a very wide range
of problems, through brute force or otherwise, I would be happy to call
them intelligent.  I suspect that most people, when faced with a machine
that could solve amazingly difficult problems, pass a Turing test, etc...,
would refer to the machine as being intelligent.  They wouldn't really care
if internally it was brute forcing stuff by running some weird quantum XYZ
system that was doing 10^10^100 calculations per second.  They
would simply see that the machine seemed to be much smarter than
themselves and thus would say it was intelligent.




for most people, that will
happen only when my system is producing results that they consider as
impressive, which will not happen soon.



Speaking of which, you're been working on NARS for 15 years!
As the theory of NARS is not all that complex (at least that was my
impression after reading you PhD thesis and a few other papers),
what's the hold up.  Even working part time I would have thought
that 15 years would have been enough to complete the system
and demonstrate its performance.

In Ben's case I understand that psynet/webmind/novamente have
all be fairly different to each other and complex.  So I understand
why it takes so long.  But NARS seems to be much simpler and
the design seems more stable over time?



It seems to me that what you are defining would be better termed
> "intelligence efficiency" rather than "intelligence".

What if I suggest to rename your notion "universal problem solver"?  ;-)



To tell the truth, I wouldn't really mind too much!  After all, once a
sufficiently powerful all purpose problem solver exists I'll simply ask
it to work out what the best way to define intelligence is and then
ask it to build a machine according to this definition.

See, even if my definition is wrong, a solution to my definition would
still succeed in solving the problem.  :-)


but I really don't see how you can put the current AGI projects, which

are as diverse one can image, into the framework you are proposing. If
you simply say that the one that don't fit in are uninteresting to
you, the others can say the same to your framework, right?



Sure, they might not want to build something that is able to achieve an
extremely wide range of goals in an extremely wide range of environments.

All I'm saying is that this is something that is very interesting to me,
and that it also seems like a pretty good definition of "intelligence".

Shane

-
This list is sponsored by AGIRI: http://www.agiri.org/email
To unsubscribe or change your options, please go to:
http://v2.listbox.com/member/?member_id=231415&user_secret=fabd7936

Re: [agi] definitions of intelligence, again?!

[Shane Legg]

Eliezer,


As the system is now solving the optimization problem in a much
> simpler way (brute force search), according to your perspective it
> has actually become less intelligent?

It has become more powerful and less intelligent, in the same way that
natural selection is very powerful and extremely stupid.



I think the best way to resolve this is to be more specific about what
we are calling powerful or stupid.

At a micro level an individual act of selection, reproduction etc. that
evolution is built upon is not powerful and is extremely stupid.

At a macro level when we consider an entire environment that performs
trillions of trillions of acts of selection, reproduction etc. over billions
of years, that system as a whole is very powerful, intelligent and creative.

The same can be said of the brain.  At a micro level an individual act of
Hebbian learning etc. on a synapse is not very powerful and quite stupid.
However, at a macro level when you consider trillions of trillions of these
acts in a system that has been trained over a couple of decades, the
result is the brain of an adult which is indeed powerful and intelligent.

Shane

-
This list is sponsored by AGIRI: http://www.agiri.org/email
To unsubscribe or change your options, please go to:
http://v2.listbox.com/member/?member_id=231415&user_secret=fabd7936

Re: [agi] definitions of intelligence, again?!

[Shane Legg]

Pei,

No. To me that is not intelligence, though it works even better.


This seems to me to be very divergent from the usual meaning
of the word intelligence.  It opens up the possibility that a super
computer that is able to win a Nobel prize by running a somewhat
efficient AI algorithm could be less intelligent than a basic pocket
calculator that is solving optimization problems in a very efficient way.

It seems to me that what you are defining would be better termed
"intelligence efficiency" rather than "intelligence".




Let is also why I think the definition of intelligence in psychology
cannot be directly accepted in AI. For human beings, problem-solving
ability at a certain age can be used approximately to indicate the
learning ability of a system (person),



I don't see this.  Unless they are designed to be highly culture neutral,
such as a Raven test, IQ tests usually test for knowledge as well as
problem solving.  In this way they can estimate how well the individual
was able to learn in the past.


They don't need to have the test in mind, indeed, but how can you

justify the authority and fairness of the testing results, if many
systems are not built to achieve what you measure?



I don't see that as a problem.  By construction universal intelligence
measures how well a system is able to act as an extremely general
purpose problem solver (roughly stated).  This is what I would like to
have, and so universal intelligence is a good measure of what I am
interested in achieving.  I happen to believe that this is also a decent
formalisation of the meaning of "intelligence" for machines.   Some
systems might be very good at what they have been designed to do,
but what I want to know is how good are they as a general purpose
problem solver?   If I can't give them a problem, by defining a goal
for them, and have them come up with a very clever solution to my
problem, they aren't what I'm interested in with my AI work.

Shane

-
This list is sponsored by AGIRI: http://www.agiri.org/email
To unsubscribe or change your options, please go to:
http://v2.listbox.com/member/?member_id=231415&user_secret=fabd7936

Re: [agi] definitions of intelligence, again?!

[Shane Legg]

Pei,

However, in general I do think that, other things being equal, the

system that uses less resources is more intelligent.



Would the following be possible with your notion of intelligence:
There is a computer system that does a reasonable job of solving
some optimization problem.  We go along and keep on plugging
more and more RAM and CPUs into the computer.  At some point
the algorithm sees that it has enough resources to always solve
the problem perfectly through brute force search and thus drops
its more efficient but less accurate search strategy.

As the system is now solving the optimization problem in a much
simpler way (brute force search), according to your perspective it
has actually become less intelligent?



NARS can...
> - accept a number as input?
> - be instructed to try to maximise this input?
> - interact with its environment in order to try to do this?
>
> I assume NARS is able to do all of these things.

Though NARS has the potential to work in the environment you
specified, it is not designed to maximize a reward measurement given
by the environment.



Of course.  If I want a general test, I can't assume that the
systems to be tested were designed with my test in mind.

Cheers
Shane

-
This list is sponsored by AGIRI: http://www.agiri.org/email
To unsubscribe or change your options, please go to:
http://v2.listbox.com/member/?member_id=231415&user_secret=fabd7936

Re: [agi] definitions of intelligence, again?!

[Shane Legg]

Pei,

Fully agree. The situation in mainstream AI is even worse on this

topic, compared to the new AGI community. Will you write something for
AGI-08 on this?



Marcus suggested that I submit something to AGI-08.  However I'm not
sure what I could submit at the moment.  I'll have a think about this after
I've finished writing my thesis in a couple of months.


if it searches different parts of the space in a context

and experience sensitive manner, it is intelligent; if it doesn't only
search among listed alternatives, but also find out new alternatives,
it is much more intelligent.



Hmmm.  Ok, imagine that you have two optimization algorithms
X and Y and they both solve some problem equally well.  The
difference is that Y uses twice as many resources as X to do it.
As I understand your notion of intelligence, X would be considered
more intelligent than Y.  True?

Essentially then, according to you intelligence depends on how well
a system can perform per unit of resources consumed?




beside input/output of
the system, you assume the rewards to be maximized come from the
environment in a numerical form, which is an assumption not widely
accepted outside the reinforcement learning community. For example,
NARS may interpret certain input as reward, and certain other input as
punishment, but it depends on many factors in the system, and is not
objective at all. For this kind of systems (I'm sure NARS isn't the
only one), how can your evaluation framework be applied?



NARS can...
- accept a number as input?
- be instructed to try to maximise this input?
- interact with its environment in order to try to do this?

I assume NARS is able to do all of these things.

Shane

-
This list is sponsored by AGIRI: http://www.agiri.org/email
To unsubscribe or change your options, please go to:
http://v2.listbox.com/member/?member_id=231415&user_secret=fabd7936

Re: [agi] definitions of intelligence, again?!

[Shane Legg]

Mark,

Gödel's theorem does not say that something is not true, but rather that
it cannot be proven to be true even though it is true.

Thus I think that the analogue of Gödel's theorem here would be something
more like:  For any formal definition of intelligence there will exist a
form of
intelligence that cannot be proven to be intelligent even though it is
intelligent.

Cheers
Shane

-
This list is sponsored by AGIRI: http://www.agiri.org/email
To unsubscribe or change your options, please go to:
http://v2.listbox.com/member/?member_id=231415&user_secret=fabd7936

Re: [agi] definitions of intelligence, again?!

[Shane Legg]

Pei,

necessary to spend some time on this issue, since the definition of

intelligence one accepts directly determines one's research goal and
criteria in evaluating other people's work. Nobody can do or even talk
about AI or AGI without an idea about what it means.



This is exactly why I am interested in the definition of intelligence.
I think the topic deserves far more time and thought in AI than it
currently gets.


Based on the above general consideration, I define "intelligence" as

"the ability to adapt and work with insufficient knowledge and
resources"





According to my definition, a thermostat is not intelligence, and nor
is an algorithm that provide "optimum" solutions by going through all
possibilities and pick the best.



If an optimisation algorithm searches some of a solution space (because
of lack of computer power to search all of it) and then returns a solution,
does this system have some intelligence according to your definition?

Most optimisation algorithms have simple adaption (let's say that it's a
genetic algorithm to make things concrete), and the system has insufficient
knowledge and resources to directly search the whole space.


To criticize his assumption as "too far away from reality" is a

different matter, which is also why I don't agree with Hutter and
Legg. Formal systems can be built on different assumptions, some of
which are closer to reality than some others.



Both AIXI and universal intelligence are too far away from reality to
be directly implemented.  I think we all agree on that.  In their current
form their main use is for theoretical study.

In the case of AIXI, it seems to me that it would be difficult to build
something that approximates these ideas in a way that produced a
real working AI system.  Or maybe I'm just not smart enough to see
how it would be done.

In the case of universal intelligence I think there is some hope due to
the fact that the C-Test is based on quite similar ideas and this has
been used to construct an intelligence test with sensible results.
Sometime after my thesis I'm going to code up an intelligence test
based on universal intelligence and see how well various AI algorithms
perform.

Cheers
Shane

-
This list is sponsored by AGIRI: http://www.agiri.org/email
To unsubscribe or change your options, please go to:
http://v2.listbox.com/member/?member_id=231415&user_secret=fabd7936

Re: [agi] Determinism

[Shane Legg]

On 5/14/07, David Clark <[EMAIL PROTECTED]> wrote:



Even though I have a Math minor from University, I have used next to no
Mathematics in my 30 year programming/design career.



Yes, but what do you program?

I've been programming for 24 years and I use math all the time.
Recently I've been working with Marcus Hutter on a new learning
algorithm based on a rather nasty mathematical derivation.  The
results kick butt.  Another tricky derivation that Hutter did a few
years back is now producing good results in processing gene
expression data for cancer research.  I could list many more...

Anyway, my point is, whether you need math in your programing
or not all depends on what it is that you are trying to program.

Shane

-
This list is sponsored by AGIRI: http://www.agiri.org/email
To unsubscribe or change your options, please go to:
http://v2.listbox.com/member/?member_id=231415&user_secret=fabd7936

Re: [agi] Tommy

[Shane Legg]

Josh,

Interesting work, and I like the nature of your approach.

We have essentially a kind of a pin ball machine at IDSIA
and some of the guys were going to work on watching this
and trying to learn simple concepts from the observations.
I don't work on it so I'm not sure what the current state of
their work is.

When you publish something on this please let the list know!

thanks
Shane

-
This list is sponsored by AGIRI: http://www.agiri.org/email
To unsubscribe or change your options, please go to:
http://v2.listbox.com/member/?member_id=231415&user_secret=fabd7936

Re: [agi] rule-based NL system

[Shane Legg]

Mark,

I was objecting to the fact that your paper did not even mention, much less

stress, the fact that prediction of previously unseen things is critical to
intelligence.



It's true that I don't talk about the prediction of "previously unseen"
things.
I suppose I just take it for granted.  I mean,  if I predictor can't
generalise
from past experience to predict patterns that have not been seen yet, but
are in some way related to the past, then it's a pretty poor predictor!

Perhaps this illustrates something more important however:  One of the
problems with defining intelligence is that it seems to involve an endless
range of abilities.  Some argue that perception is very important.  Some
that
language is key.  Others think that mathematical thinking, or probabilistic
reasoning, or rational thought are very important.  Perhaps planning is
very important. Maybe the ability to abstract.  Emotions?  And so on...

The beauty of the approach that we have taken is that we abstract above
all this and just talk in terms of goal achieving performance.  Thus things
such as planning, reasoning or prediction are important to the extent that
they enable the agent to achieve goals.  If they don't enable the system
to work better in some environment, then they aren't a part of intelligence.
Not only does this allow us to not have to name what each of these things
are, it also means that the intelligence test measures cognitive abilities
that may enhance an agent's performance that we have not yet thought of.


I agree with the first sentence of your second paragraph entirely but point

out that a "by rote" machine with virtually infinite experience will test as
if it had high universal intelligence unless the test manages to hit upon
some area where it didn't have experience -- and I feel that this is
entirely incorrect.



Actually this is another nice feature of universal intelligence: An agent
can't
really get a better score by having a large database of prior knowledge.

The problem is that if you put some information into the agent's database,
"the grass is green and the sky is blue" then this will help the agent in a
world where this is true.  However, there will be another world with about
the
same complexity where the sky is green and the grass blue.  The agent has
no idea which of these two worlds it is going to face.  This makes a
database
of prior experience is useless.  The only way that the agent can do well is
by
quickly learning from its experience and adapting to deal with the
uncertainties
in its environment.

Unfortunately, in the 8 page Benelearn paper there wasn't space to get into
many of these interesting aspects of this intelligence test.

Cheers
Shane

-
This list is sponsored by AGIRI: http://www.agiri.org/email
To unsubscribe or change your options, please go to:
http://v2.listbox.com/member/?member_id=231415&user_secret=fabd7936

Re: [agi] rule-based NL system

[Shane Legg]

On 5/2/07, Mark Waser <[EMAIL PROTECTED]> wrote:



One of the things that I think is *absolutely wrong* about Legg's
paper is that he only uses more history as an example of generalization.  I
think that predictive power is test for intelligence (just as he states) but
that it *must* include things that the agent has never seen before.  In this
sense, I think that Legg's paper is off the mark to the extent of being
nearly useless (since you can see how it's has poisoned poor Matt's
approach).



Mark,

Why do you think that a Legg-Hutter style intelligence test would
not expose an agent to things it hadn't seen before?

To have a significant level of intelligence an agent must be able
to deal with environments that are full of surprises and unknowns.
Agents that can't do this would only be able to deal with the most
basic environments, and thus would have a relatively low universal
intelligence value.

Cheers
Shane

-
This list is sponsored by AGIRI: http://www.agiri.org/email
To unsubscribe or change your options, please go to:
http://v2.listbox.com/member/?member_id=231415&user_secret=fabd7936

Re: [agi] Circular definitions of intelligence

[Shane Legg]

Ben,

Are you claiming that the choice of "compiler constant" is not pragmatically


significant in the definition of the Solomonoff-Levin universal prior, and
in Kolmogorov
complexity?  For finite binary sequences...

I really don't see this, so it would be great if you could elaborate.



In some cases it matters, in others it doesn't.  Solomonoff's prediction
error
theorem shows that the total summed expected squared prediction error is
bounded by a constant when the true generating distribution \mu is
computable.
The constant is (ln 2)/2 K(\mu) bits.  The K term in this bound depends on
the
choice of reference machine.  For a reasonable choice of reference machine
you might be able to push the bound up by something like 1000 bits.  If you
are considering long running systems that will process large amounts of
data,
that 1000 extra bits is tiny.

On the other hand, if you want to know if K(10) < K(147) then your answer
will depend on which reference machine you use.  In short: Kolmogorov
complexity works well for reasonably big objects, it doesn't work well for
small objects.

Probably the best solution is to condition the measure with information
about
the world.  In which case K(10|lots of world data) < K(147|lots of world
data)
should work the way you expect.  Google complexity works this way.
In the case of Solomonoff induction, you let the predictor watch the world
for
a while before you start trying to get it to solve prediction tasks.


In a practical Novamente context, it seems to make a big difference.  If we

make different choices regarding the internal procedure-representation
language Novamente uses, this will make a big difference in what
internally-generated programs NM thinks are simpler ... which will make a
big difference in which ones it retains versus forgets; and which ones it
focuses its attention on and prioritizes for generating actions.



I think that the universal nature we see in Kolmogorov complexity should
also apply to practical AGI systems.  By that I mean the following:

By construction, things which have high Kolmogorov complexity are complex
with respect to any reasonable representation system.  In essence, the
reference
machine is your representation system.  Once an AGI system has spent some
time learning about the world I expect that it will also find that there are
certain
types of representation systems that work well for certain kinds of
problems.
For example, it might encounter a problem that seems complex, but then it
realises that, say, if it views the problem as a certain type of algebra
problem
then it knows how to find a solution quite easily.  I think that the hardest
part
to finding a solution to a difficult problem often lies in finding the right
way to
view the problem, in order words, the right representation.

Cheers
Shane

-
This list is sponsored by AGIRI: http://www.agiri.org/email
To unsubscribe or change your options, please go to:
http://v2.listbox.com/member/?member_id=231415&user_secret=fabd7936

Re: [agi] Circular definitions of intelligence

[Shane Legg]

Mike,

But interestingly while you deny that the given conception of intelligence

is rational and deterministic.. you then proceed to argue rationally and
deterministically.



Universal intelligence is not based on a definition of what rationality is.
It is based
on the idea of achievement.  I believe that if you start to behave
irrationally (by any
reasonable definition of the word) then your ability to achieve goals will
go down and
thus so will your universal intelligence.


that actually you DON'T usually know what you desire. You have conflicting

desires and goals. [Just how much do you want sex right now? Can you produce
a computable function for your desire?]



Not quite.  Universal intelligence does not require that you personally can
define
your, or some other system's, goal.  It just requires that the goal is well
defined
in the sense that a clear definition could be written down, even if you
don't know
what that would look like.

If you want intelligence to include undefinable goals in the above weaker
sense
then you have this problem:

"Machine C is not intelligent because it cannot do X, where X is something
that cannot be defined."

I guess that this isn't a road you want to take as I presume that you think
that
machine intelligence is possible.


And you have to commit yourself at a given point, but that and your

priorities can change the next minute.



A changing goal is still a goal, and as such is already taken care of by the
universal intelligence measure.


And vis-a-vis universal intelligence, I'll go with Ben


"According to Ben Goertzel, Ph. D, "Since universal intelligence is only
definable up to an arbitrary constant, it's of at best ~heuristic~ value in
thinking about the constructure of real AI systems. In reality, different
universally intelligent modules may be practically applicable to different
types of problems." [8] "



Ben's comment is about AIXI, so I'll change to that for a moment.  I'm going
to have
to be a bit more technical here.

I think the compiler constant issue with Kolmogorov complexity is in some
cases
important, and in others it is not.  In the case of Solomonoff's continuous
universal
prior (see my Scholarpedia article on algorithmic probability theory for
details) the
measure converges to the true measure very quickly for any reasonable choice
of
reference machine.  With different choices of reference machine the compiler
constant may mean that the system doesn't converge for a few more bytes of
input.
This isn't an issue for an AGI system that will be processing huge amounts
of data
over time.  The optimality of its behaviour in the first hundred bytes of
its existence
really doesn't matter.  Even incomputable super AIs go through an infantile
stage,
albeit a very short one.


You seem to want to pin AG intelligence down precisely, I want to be more

pluralistic - and recognize that uncertainty and conflict are fundamental to
its operation.



Yes, I would like to pin intelligence down as precisely as possible.

I think that if somebody could do this it would be a great step forward.
I believe that issues of definition and measurement are the bedrock of
good science.

Cheers
Shane

-
This list is sponsored by AGIRI: http://www.agiri.org/email
To unsubscribe or change your options, please go to:
http://v2.listbox.com/member/?member_id=231415&user_secret=fabd7936

Re: [agi] mouse uploading

[Shane Legg]

Numbers for humans vary rather a lot.  Some types of cells have up to
200,000 connections (Purkinje neurons) while others have very few.
Thus talking about "the" number of synapses per neuron doesn't make
much sense.  It all depends on which type of neuron etc. you mean.

Anyway, when talking about a global brain average I most often see the
number 1,000.   For rat cortex (which is a bit different to mouse cortex
in terms of thickness and density) I usually see the number 10,000 as
the average (just for cortex, not the whole brain).

Shane


On 4/29/07, Matt Mahoney <[EMAIL PROTECTED]> wrote:


Does anyone know if the number of synapses per neuron (8000) for mouse
cortical cells also apply to humans?  This is the first time I have seen
an
estimate of this number.  I believe the researchers based their mouse
simulation on anatomical studies.

--- "J. Storrs Hall, PhD." <[EMAIL PROTECTED]> wrote:

> In case anyone is interested, some folks at IBM Almaden have run a
> one-hemisphere mouse-brain simulation at the neuron level on a Blue Gene
(in
>
> 0.1 real time):
>
> http://news.bbc.co.uk/2/hi/technology/6600965.stm
> http://ieet.org/index.php/IEET/more/cascio20070425/
> http://www.modha.org/papers/rj10404.pdf which reads in gist:
>
> Neurobiologically realistic, large-scale cortical and sub-cortical
> simulations
> are bound to play a key role in computational neuroscience and its
> applications to cognitive computing. One hemisphere of the mouse cortex
has
> roughly 8,000,000 neurons and 8,000 synapses per neuron. Modeling at
this
> scale imposes tremendous constraints on computation, communication, and
> memory capacity of any computing platform.
>  We have designed and implemented a massively parallel cortical
simulator
> with
> (a) phenomenological spiking neuron models; (b) spike-timing dependent
> plasticity; and (c) axonal delays.
>  We deployed the simulator on a 4096-processor BlueGene/L supercomputer
with
>
> 256 MB per CPU. We were able to represent 8,000,000 neurons (80%
excitatory)
>
> and 6,300 synapses per neuron in the 1 TB main memory of the system.
Using a
>
> synthetic pattern of neuronal interconnections, at a 1 ms resolution and
an
> average firing rate of 1 Hz, we were able to run 1s of model time in 10s
of
> real time!
>
> Josh
>
> -
> This list is sponsored by AGIRI: http://www.agiri.org/email
> To unsubscribe or change your options, please go to:
> http://v2.listbox.com/member/?&;
>


-- Matt Mahoney, [EMAIL PROTECTED]

-
This list is sponsored by AGIRI: http://www.agiri.org/email
To unsubscribe or change your options, please go to:
http://v2.listbox.com/member/?&;



-
This list is sponsored by AGIRI: http://www.agiri.org/email
To unsubscribe or change your options, please go to:
http://v2.listbox.com/member/?member_id=231415&user_secret=fabd7936

Re: [agi] Circular definitions of intelligence

[Shane Legg]

Mike,

1) It seems to assume that intelligence is based on a rational,

deterministic program - is that right? Adaptive intelligence, I would argue,
definitely isn't. There isn't a rational, right way to approach the problems
adaptive intelligence has to deal with.



I'm not sure what you mean by this.  The agent that we measure the
intelligence of does
not have to be deterministic, nor does the environment.  Indeed, the agent
doesn't even
have to have a computable probability distribution, it could work by magic
for all we care.


2) It assumes that intelligent agents maximise their rewards. Wrong. You

don't except in extreme situations try to maximise your rewards when you
invest on the stockmarket - or invest in any other action.

In the real world, you have to decide how to invest your time, energy and
resources in taking/solving problematic decisions/problems (like how to
invest on the stockmarket). Those decisions carry rewards, risks and
uncertainty.  The higher the rewards, the higher the risks (nor just of
failure but of all kinds of danger). The lower the rewards, the lower the
risks (and the greater the security).



Let's say that you want to invest money in a low risk way that still has
some minimal
level of return.  In other words, you don't want to simply maximise your
expected
return, rather you want to maximise some balance of return and risk (or any
other
things you also want to take into account such as time and energy).  Take
all these
factors and define a utility function over the possible outcomes.  If you
can't do this
then you don't really know exactly what it is that you desire.  Now simply
consider
the reward signal from the environment to the agent to be exactly the
utility function
that you just defined.   In order to perform well in this setting the agent
must work
out how to balance return on investment against risks etc.  Moreover, this
type of
environment still has a computable measure and thus is already contained in
our
intelligence test.




3) And finally, just to really screw up this search for intelligence
definitions - any definition will be fundamentally ARBITRARY There will
always be conflicting ideals of what intelligent problem solving involves..



There is no such thing as a provably true definition.  However some
definitions are
clearer, more general and more consistent with the informal usage than
others.

So let me put the challenge to you: Can you name one well defined process to
do
with intelligent problem solving that universal intelligence doesn't already
test for?

Shane

-
This list is sponsored by AGIRI: http://www.agiri.org/email
To unsubscribe or change your options, please go to:
http://v2.listbox.com/member/?member_id=231415&user_secret=fabd7936

Re: [agi] Circular definitions of intelligence

[Shane Legg]

Kaj,

(Disclaimer: I do not claim to know the sort of maths that Ben and

Hutter and others have used in defining intelligence. I'm fully aware
that I'm dabbling in areas that I have little education in, and might
be making a complete fool of myself. Nonetheless...)



I'm currently writing my PhD thesis at the moment in which, at Hutter's
request, I am going to provide what should be an easy to understand
explanation of AIXI and the universal intelligence measure.  Hopefully
this will help make the subject more understandable to people outside
the area of complexity theory.  I'll let this list know when this is out.


The intelligence of a system is a function of the amount of different

arbitrary goals (functions that the system maximizes as it changes
over time) it can carry out and the degree by which it can succeed in
those different goals (how much it manages to maximize the functions
in question) in different environments as compared to other systems.



This is essentially what Hutter and I do.  We measure the performance
of the system for a given environment (which includes the goal) and
then sum them up.  The only additional thing is that we weight them
according to the complexity of each environment.  We use Kolmogorov
complexity, but you could replace this with another complexity measure
to get a computable intelligence measure.  See for example the work of
Hernandez (which I reference in my papers on this).  Once I've finished
my thesis, one thing that I plan to do is to write a program to test the
universal intelligence of agents.


This would eliminate a thermostat from being an intelligent system,

since a thermostat only carries out one goal.



Not really, it just means that the thermostat has an intelligence of one
on your scale.  I see no problem with this.  In my opinion the important
thing is that an intelligence measure orders things correct.  For example,
a thermostat should be more intelligent than a system that does nothing.
A small machine learning algorithm should be smarter still, a mouse
smarter still, and so on...


Humans would be

classified as relatively intelligent, since they can be given a wide
variety of goals to achieve. It also has the benefit of assigning
narrow-AI systems a very low intelligence, which is what we want it to
do.



Agreed.

If you want to read about the intelligence measure that I have developed
with Hutter check out the following.
A summary set of talk slides:

http://www.vetta.org/documents/Benelearn-UniversalIntelligence-Talk.pdf

Or for a longer paper:

http://www.vetta.org/documents/ui_benelearn.pdf

Unfortunately the full length journal paper (50 pages) is still in review so
I'm not sure when that will come out.  But my PhD thesis will contain this
material and that should be ready in a few months time.

Cheers
Shane

-
This list is sponsored by AGIRI: http://www.agiri.org/email
To unsubscribe or change your options, please go to:
http://v2.listbox.com/member/?member_id=231415&user_secret=fabd7936

Re: [agi] Growing a Brain in Switzerland

[Shane Legg]

On 4/5/07, Eugen Leitl <[EMAIL PROTECTED]> wrote:


>I forget the exact number, but I think something like 20% of the
human
>genome describes the brain.  If somebody is interested in building a

No, it codes for the brain tissue. That's something very different from
describing the brain. See



I didn't mean to imply that all this was for wiring, just that there is a
sizable
about of information used to construct the brain that comes from the genes.
If you want to model the brain then this is the kind of information that you
are going to have to put into your model.

Why does the optic tract project to the lateral geniculate nucleus, the
pretectum
and the superior colliculus and not other places in the brain?  Why does the
lateral genicultate body project to striate and not other parts of cortex?
Why does
the magnocellular pathway project to layer 4Calpha, while the parvocullular
pathway projects to 4A and 4Cbeta?  Why does the cerebral cortex project to
the putamen and caudate nucleus, but not the subthalamic nucleus?  I could
list pages and pages of examples of brain wiring that you were born with and
that came from your genetics, it's basic neuro science.

I don't clam that all wiring in the brain is genetic, or even a sizable
proportion of it.
What I am claiming is that the brain wiring that is genetic is non-trivial
and cannot
be ignored if somebody wants to build a working brain simulation.


You remember the thread: complexity in the code versus complexity in the

data? The Blue Brain complexity is all in the data. This is very different
from the classical AI, which tends to obsessionate about lots of clever
algorithms, but typically does sweep the data (state) under the carpet.



Yes, I agree, it's in the "data" rather than the "code".  But I don't accept
that you can say that their model is simple.

Shane

-
This list is sponsored by AGIRI: http://www.agiri.org/email
To unsubscribe or change your options, please go to:
http://v2.listbox.com/member/?list_id=303

Re: [agi] Growing a Brain in Switzerland

[Shane Legg]

On 4/4/07, Eugen Leitl <[EMAIL PROTECTED]> wrote:


how do you reconcile the fact that babies are very stupid compared to
> adults?  Babies have no less genetic hardware than adults but the
difference


The wiring is not determined by the genome, it's only a facility envelope.


Some wiring is genetic, and some is not.  On a large scale genes regulate
how one part of the neocortex is wired to another part (there are even
little
tiny crawler things that do the wiring up during the prenatal development of
the brain that sound totally science fiction and very cool, though the
system
isn't exactly perfect as they kill quite a few neurons when they try to craw
around the brain hooking all the wiring up).

At a micro scale each of the different types of neurons have different
dendritic tree structures (which is genetic), and lie in particular layers
of
cortex (which is also genetic), and various other things.  In short, it's
not
really genetic or due to adaption, it's a complex mixture of genetics and
adaption that produces the wiring in the adult brain.


The models are not complex. The emulation part is a standard numerics

package.



Heh.  Come to Switzerland and talk to the Blue Brain guys at EPFL...
Their model is very complex and definitely not simply some standard
numerics package.  They are working in collaboration with something
like 400 researchers all around the world and the project will be going
for at least several decades.  Simple it is not.

Shane

-
This list is sponsored by AGIRI: http://www.agiri.org/email
To unsubscribe or change your options, please go to:
http://v2.listbox.com/member/?list_id=303

Re: [agi] Growing a Brain in Switzerland

[Shane Legg]

On 4/5/07, Eugen Leitl <[EMAIL PROTECTED]> wrote:


On Thu, Apr 05, 2007 at 11:34:09AM +0200, Shane Legg wrote:

>  The wiring is not determined by the genome, it's only a facility
>  envelope.
>
>Some wiring is genetic, and some is not.  On a large scale genes

No, the wiring is not genetically determined. There is no specific gene
telling
cell 0x093785 to connect to cell 0x3464653, and a similiar gene for every
other



That two specific neurons are not wired together due to genetics does not
mean
that there is not wiring that is genetically determined.  The brain contains
a huge
about of wiring information that comes from the genes.




I've never claimed anything otherwise. But the genome is not a noticeable
source of complexity in the adult individual.



I forget the exact number, but I think something like 20% of the human
genome describes the brain.  If somebody is interested in building a model
of a working brain that's a lot of complexity.



   numerics package.  They are working in collaboration with something
>like 400 researchers all around the world and the project will be
>going for at least several decades.  Simple it is not.

It is not only simple, it is completely trivial in comparison to
a hypothetical AI designed by people. The complexity in the behaviour
comes from the neuroanatomy, not the code.



What you said was, "the models are not complex".

What they are modeling IS the neuroanatomy and its behaviour.

Shane

-
This list is sponsored by AGIRI: http://www.agiri.org/email
To unsubscribe or change your options, please go to:
http://v2.listbox.com/member/?list_id=303

Re: [agi] My proposal for an AGI agenda

[Shane Legg]

On 3/23/07, David Clark <[EMAIL PROTECTED]> wrote:



Both the code and algorythmn must be good for any computer system to work
and neither is easy.  The bond formula was published for many years but this
particular company certainly didn't have a copy of it inside a program they
could use.  The formula was 1 line of at least 4K lines of code.  The
program wasn't so trivial either :)



The reason AGI doesn't exist yet is not because there aren't enough skilled
programmers in the world.  (Or that these people are using the wrong
operating
system or programming language etc... to address the rest of the discussion
on this thread!)  The problem is that people aren't exactly clear about what
it is
that has to be programmed.  Time and time again in the field people have
thought
that they knew what had to be done, and yet when they finally got around to
coding
it the results weren't what they had hoped for.




Is the research on AI full of Math because there are many Math professors
that publish in the field or is the problem really Math related?  Many PhDs
in computer science are Math oriented exactly because the professors that
deem their work worth a PhD are either Mathematicians or their sponsoring
professor was.



I don't know of any math profs who publish in artificial intelligence,
though no doubt
there are a few that do.  No, thinking about it now I can think of a few.
Even if you
look at the work of my PhD supervisor Marcus Hutter, he's not a math prof,
he's
actually a physics PhD.  His work might look very mathematical to a computer
scientist, but he doesn't actually use much beyond about 4th year university
level
mathematics and statistics in his book.  Indeed he likes to joke about
mathematicians
and how they are overly formal and concerned about details like whether he
has
properly deal with certain special cases on sets of measure 0 :-)

So yeah, the math that you see in the field is almost always coming from
people
who are mathematically inclined, but aren't math profs.  I should also note
that the
number of pure theory people in AI is very small.  For example, I went to
the ALT
conference last year to present some work and there were only 40 people.
This is
the second biggest AI theory conference in the world (after COLT).  Other
areas like
genetic algorithms attract thousands.

Shane

-
This list is sponsored by AGIRI: http://www.agiri.org/email
To unsubscribe or change your options, please go to:
http://v2.listbox.com/member/?list_id=303

Re: [agi] My proposal for an AGI agenda

[Shane Legg]

On 3/23/07, David Clark <[EMAIL PROTECTED]> wrote:


I have a Math minor from University but in 32 years of computer work, I
haven't used more than grade 12 Math in any computer project yet.


...


I created a bond comparison program for a major wealth investment firm
that
used a pretty fancy formula at it's core but I just typed it in.  I didn't
have to create it, prove it or even understand exactly why it was any
good.



IMO, creating an AGI isn't really a programming problem.  The hard part is
knowing exactly what to program.  The same was probably true of your bond
program: The really hard part was originally coming up with that 'fancy
formula'
which you just had to type in.

Thus far math has proven very useful in many areas of artificial
intelligence,
just pick up any book on machine learning such as Bishop's.  Whether it will
also be of large use for AGI... only time will tell.

Shane

-
This list is sponsored by AGIRI: http://www.agiri.org/email
To unsubscribe or change your options, please go to:
http://v2.listbox.com/member/?list_id=303

Re: [agi] My proposal for an AGI agenda

[Shane Legg]

On 3/21/07, Chuck Esterbrook <[EMAIL PROTECTED]> wrote:

Sometimes the slowness of a program is not contained in a small

portion of a program.



Sure.  For us however this isn't the case.

Cobra looks nice, very clean to read, even more so than Python.
However the fact that it's in beta and .NET sort of kills it for us.

As we will be going into high performance computing, we have
no choice but to do the core work in plain C running on Linux.

Shane

-
This list is sponsored by AGIRI: http://www.agiri.org/email
To unsubscribe or change your options, please go to:
http://v2.listbox.com/member/?list_id=303

Re: [agi] My proposal for an AGI agenda

[Shane Legg]

Ben, I didn't know you were a Ruby fan...

After working in C# with Peter I'd say that's is a pretty good choice.
Sort of like Java but you can get closer to the metal where needed
quite easily.

For my project we are using Ruby and C.  Almost all the code can
be in high level Ruby which is very fast to code and modify, and then
the few parts of the code that consume 99.9% of the CPU time get
converted into C and carefully optimised.

Shane

-
This list is sponsored by AGIRI: http://www.agiri.org/email
To unsubscribe or change your options, please go to:
http://v2.listbox.com/member/?list_id=303

Re: [agi] one-shot Turing test

[Shane Legg]

On 3/9/07, J. Storrs Hall, PhD. <[EMAIL PROTECTED]> wrote:


Perhaps the ultimate Turing Test would be to make the system itself act as
the
interviewer for a Turing Test of another system.



It's called an "inverted Turing test". See:

Watt, S. (1996) Naive-Psychology and the Inverted Turing Test. Psycoloquy
7(14)


Shane

-
This list is sponsored by AGIRI: http://www.agiri.org/email
To unsubscribe or change your options, please go to:
http://v2.listbox.com/member/?list_id=303

Re: [agi] Numenta (Hawkins) software released

[Shane Legg]

The second scary bit, which I didn't mention above, is made clear in the
blog post from the company CEO, Donna Dubinsky:


Why do we offer you a license without deployment rights? Well, although we
are very excited about the ultimate applications of this technology, we feel
it is too early to expect commercial class applications. We do not want to
mislead you as to the state of the technology.

Perhaps we're wrong! If you find yourself closer to a commercial
application than we expect, let us know. We promise that we'll speed up our
commercial licensing plans!!


which is from http://www.numenta.com/for-developers/blog.php

In other words: Go make cool stuff with our technology, however if you are
about
to make lots of money, please let us know and we will then calculate how
much
it is going to cost you for a licence.

There is no way I'd ever agree to terms like that, especially given that
they cover not
just code, but also all business ideas etc. related to this technology.

Shane

-
This list is sponsored by AGIRI: http://www.agiri.org/email
To unsubscribe or change your options, please go to:
http://v2.listbox.com/member/?list_id=303

Re: [agi] Numenta (Hawkins) software released

[Shane Legg]

It might however be worth thinking about the licence:

Confidentiality. 1. Protection of Confidential Information. You agree that

all code, inventions, algorithms, business concepts, workflow, ideas, and
all other business, technical and financial information, including but not
limited to the HTM Algorithms, HTM Algorithms Source Code, and HTM
Technology, that you obtain or learn from Numenta in connection with this
Agreement are the confidential property of Numenta (Confidential
Information). Except as authorized herein, you will hold in confidence and
not use, except as permitted or required in the Agreement, or disclose any
Confidential Information and you will similarly bind your employees in
writing. You will not be obligated under this Section 6 with respect to
information that you can document: (i) is or has become readily publicly
available without restriction through no fault of you or your employees or
agents; or (ii) is received without restriction from a third party lawfully
in possession of such information and lawfully empowered to disclose such
information; or (iii) was rightfully in your possession without restriction
prior to its disclosure by Numenta; or (iv) was independently developed by
your employees or consultants without access to such Confidential
Information.



Shane


On 3/7/07, J. Storrs Hall, PhD. <[EMAIL PROTECTED]> wrote:


Just noticed this on Slashdot.
Open source but not free software, for those of you for whom this makes a
difference.
http://www.numenta.com/for-developers/software.php

Josh

-
This list is sponsored by AGIRI: http://www.agiri.org/email
To unsubscribe or change your options, please go to:
http://v2.listbox.com/member/?list_id=303



-
This list is sponsored by AGIRI: http://www.agiri.org/email
To unsubscribe or change your options, please go to:
http://v2.listbox.com/member/?list_id=303

Re: [agi] Has anyone read "On Intelligence"

[Shane Legg]

Sorry, the new version of the book I mentioned (I read the old one) is
called "Principles of Neural Science".

With regards to computer power, I think it is very important.  The average
person doing research in AI (i.e. a PhD grad student) doesn't have access
to much more than a PC or perhaps a small cluster of PCs.  So it's all
very well that IBM can build super computers with vast amounts of power,
but most of us don't get access to such machines --- we're many orders
of magnitude behind this.

The other thing is that what is necessary for some algorithm to solve
a problem is very different to what was needed to develop the algorithm
in the first place.  To develop a machine learning algorithm you might want
to test it on 10 different data sets, with various different parameter
settings,
and a few different versions of the algorithm, and then run it many times in
each of these configurations order to get accurate performance statistics.
Then you look at the results and come up with some new ideas and repeat.
Thus, even if algorithm Y is a decent AGI when run on hardware X, you
probably want 100X computer power in order to develop algorithm Y.

Shane

-
This list is sponsored by AGIRI: http://www.agiri.org/email
To unsubscribe or change your options, please go to:
http://v2.listbox.com/member/?list_id=303

Re: [agi] Has anyone read "On Intelligence"

[Shane Legg]

I think "On Intelligence" is a good book.  It made an impact on
me when I first read it, and it lead to me reading a lot more neuro
science since then.  Indeed in hindsight is seems strange to me
that I was so interested in AGI and yet I hadn't seriously studied
what is known about how the brain works.  Indeed almost nobody
in AI does, even people working on artificial neural networks.

Anyway, having read a fair amount of neuro science since then
it has become clear to me that while Hawkins' book is a good
understandable summary of a particular view of neuro science,
the claims he makes are all either well known facts, or things
which a fair number of neuro scientists already believe.  So there
isn't anything really new in there that I know of.

The other thing is that he presents a greatly simplified view of how
things really work.  This makes the book readable for non-scientists,
which is great, however nobody really knows how much of all those
details he glosses over are unimportant implementation stuff, and how
much of it is critical to understanding how the whole system behaves.
Of course, nobody will really know this for sure until the brain is fully
understood.

If you're read On Intelligence and are interested in a basic undergraduate
overview of neuroscience I'd recommend the classic text book "Essentials
of Neural Science and Behavior" by Kandel, Schwartz and Jessell.  Once
you've read that much of the scientific literature in the field is
understandable.

Shane



On 2/21/07, Aki Iskandar <[EMAIL PROTECTED]> wrote:


I'd be interested in getting some feedback on the book "On
Intelligence" (author: Jeff Hawkins).

It is very well written - geared for the general masses of course -
so it's not written like a research paper, although it has the feel
of a thesis.

The basic premise of the book, if I can even attempt to summarize it
in two statements (I wouldn't be doing it justice though) is:

1 - Intelligence is the ability to make predictions on memory.
2 - Artificial Intelligence will not be achieved by todays computer
chips and smart software.  What is needed is a new type of computer -
one that is physically wired differently.


I like the first statement.  It's very concise, while capturing a
great deal of meaning, and I can relate to it ... it "jives".

However, (and although Hawkins backs up the statements fairly
convincingly) I don't like the second set of statements.  As a
software architect (previously at Microsoft, and currently at Charles
Schwab where I am writing a custom business engine, and workflow
system) it scares me.   It scares me because, although I have no
formal training in AI / Cognitive Science, I love the AI field, and
am hoping that the AI puzzle is "solvable" by software.

So - really, I'm looking for some of your gut feelings as to whether
there is validity in what Hawkins is saying (I'm sure there is
because there are probably many ways to solve these type of
challenges), but also as to whether the solution(s) its going to be
more hardware - or software.

Thanks,
~Aki

P.S.  I remember a video I saw, where Dr. Sam Adams from IBM stated
"Hardware is not the issue.  We have all the hardware we need".
This makes sense.  Processing power is incredible.  But after reading
Hawkins' book, is it the right kind of hardware to begin with?

-
This list is sponsored by AGIRI: http://www.agiri.org/email
To unsubscribe or change your options, please go to:
http://v2.listbox.com/member/?list_id=303



-
This list is sponsored by AGIRI: http://www.agiri.org/email
To unsubscribe or change your options, please go to:
http://v2.listbox.com/member/?list_id=303

Re: [agi] Why so few AGI projects?

[Shane Legg]

Eliezer,Shane, what would you do if you had your headway?  Say, you won the
lottery tomorrow (ignoring the fact that no rational person would buy aticket).  Not just "AGI" - what specifically would you sit down and doall day?I've got a list of things I'd like to be working on.  For example, I'd like to
try to build a universal test of machine intelligence, I've also got ideas inthe area of genetic algorithms, neural network architectures, and somemore theoretical things related to complexity theory and AI.  I also want to
spend more time learning neuroscience.  I think my best shot at buildingan AGI will involve bringing ideas from many of these areas together.
Indeed not.  It takes your first five years simply to figure out whichway is up.  But Shane, if you restrict yourself to results you canregularly publish, you couldn't work on what you really wanted to do,even if you had a million dollars.
If I had a million dollars I wouldn't care so much about my "career"as I wouldn't be dependent on the academic system to pay my bills.As such I'd only publish once, or perhaps twice, a year and would
spend more time on areas of research that were more likely to failor would require large time investments before seeing results.Shane

This list is sponsored by AGIRI: http://www.agiri.org/email
To unsubscribe or change your options, please go to:
http://v2.listbox.com/member/[EMAIL PROTECTED]

Re: [agi] Why so few AGI projects?

[Shane Legg]

This is a question that I've thought about from time to time.  The conclusionI've come to is that there isn't really one or two reasons, there are many.Surprisingly, most people in academic AI aren't really all that into AI.
It's a job.  It's more interesting than doing database programming ina bank, but at the end of the day it's just a job.  They're not out tochange the world or do anything amazing, it's hard enough just trying
to get a paper into conference X or Y.  It's true that they are skepticalabout whether AI will make large progress towards human levelintelligence in their life times, however I think the more important pointis that they simply don't even think about this question.  They're just not
interested.  I'd say that this is about 19 out of every 20 people in academicAI.  Of course there are thousands of people working in academic AIaround the world, so 1 out of 19 is still a sizable number of people in total.
Funding is certainly a problem.  I'd like to work on my own AGI ideasafter my PhD is over next year... but can I get money to do that?  Probablynot.  So as a compromise I'll have to work on something else in AI during
the day, and spend my weekends doing the stuff I'd really like to be doing.Currently I code my AI at nights and weekends.Pressure to publish is also a problem.  I need results on a regular basisthat I can publish otherwise my career is over.  AGI is not really short term
results friendly.Another thing is visibility.  Of the academic people I know who are tryingto build a general artificial intelligence (although probably not saying quitethat in their papers), I would be surprised if any of them were known to
anybody on this list.  These a non-famous young researchers, and becausethey can't publish papers saying that they want build a thinking machine,you'd only know this if you were to meet them in person.
One thing that people who are not involved in academic AI often don'tappreciate is just how fractured the field is.  I've seen plenty of exampleswhere there are two sub-fields that are doing almost the same thing
but which are using different words for things, go to different conferences,and cite different sets of people.  I bring this up because I sometimesget the feeling that some people think that "academic AI" is some sort
of definable group.  In reality, most academics lack of knowledge aboutAGI is no different to their lack of knowledge of many other areas of AI.In other words, they aren't ignoring AGI any more than they are ignoring
twenty other areas in the field.Shane

This list is sponsored by AGIRI: http://www.agiri.org/email
To unsubscribe or change your options, please go to:
http://v2.listbox.com/member/[EMAIL PROTECTED]

Re: [agi] Definitions of intelligence

[Shane Legg]

Pei, I'll correct your definition and add the new ones youcite on Monday.ThanksShaneOn 9/1/06, Pei Wang <
[EMAIL PROTECTED]> wrote:Shane,Thanks for the great job! It will be a useful resource for all of us.
In my definition, I didn't use the word "agent", but "system".You may also want to consider the 8 definitions listed in AIMA(http://aima.cs.berkeley.edu/
), page 2.PeiOn 9/1/06, Shane Legg <[EMAIL PROTECTED]> wrote:>> As part of some research I've been doing with Prof. Hutter on> AIXI and formal definitions of machine intelligence, I've been
> collecting definitions of intelligence that have been proposed> by psychologists and AI researchers.  I now have about 60> definitions.  At least to the best of my knowledge, this makes> it the largest and most fully referenced collection that there is.
>> If you're interested I've put it online:>>  http://www.idsia.ch/~shane/intelligence.html>> Of course to really understand the definitions you will need
> to follow the references back to the sources.  Nevertheless,> this collection provides a compact overview.>> Corrects etc. are of course welcomed, as are new definitions,> provided that they have been published somewhere so that
> I can properly cite them.  Also the individual must be of a> certain significance the area, e.g. be a psychologist, or an> academic AI researcher, run an AI company etc.>> Shane>
>  >  To unsubscribe, change your address, or temporarily deactivate your> subscription, please go to> 
http://v2.listbox.com/member/[EMAIL PROTECTED]---To unsubscribe, change your address, or temporarily deactivate your subscription,please go to 
http://v2.listbox.com/member/[EMAIL PROTECTED]

To unsubscribe, change your address, or temporarily deactivate your subscription, 
please go to http://v2.listbox.com/member/[EMAIL PROTECTED]

[agi] Definitions of intelligence

[Shane Legg]

As part of some research I've been doing with Prof. Hutter onAIXI and formal definitions of machine intelligence, I've beencollecting definitions of intelligence that have been proposedby psychologists and AI researchers.  I now have about 60
definitions.  At least to the best of my knowledge, this makesit the largest and most fully referenced collection that there is.If you're interested I've put it online:
http://www.idsia.ch/~shane/intelligence.htmlOf course to really understand the definitions you will needto follow the references back to the sources.  Nevertheless,this collection provides a compact overview.
Corrects etc. are of course welcomed, as are new definitions,provided that they have been published somewhere so thatI can properly cite them.  Also the individual must be of acertain significance the area, 
e.g. be a psychologist, or anacademic AI researcher, run an AI company etc.Shane

To unsubscribe, change your address, or temporarily deactivate your subscription, 
please go to http://v2.listbox.com/member/[EMAIL PROTECTED]

Re: [agi] Marcus Hutter's lossless compression of human knowledge prize

[Shane Legg]

On 8/13/06, Pei Wang <[EMAIL PROTECTED]> wrote:
Shane,I don't think the statement "the universe is computable" can either beproved or disproved, because it is not a mathematical statement.However, there can be evidence for or against it.
Agreed.  To date there is no evidence against it, and the entire fieldof standard physics for it.  That's a lot of evidence on one side, andno evidence at all on the other.  Thus it is pretty reasonable to make
the assumption that the environment is computable when doing thiskind of work.Secondly, while one cannot prove that the whole universe is computable,the observed universe to date has proven to be computable to a very high
degree.  Thus, even if this assumption was violated, it would still hold formuch of the universe and to a very high degree of accuracy.  Clearly thenany AI based on the computable universe assumption would still continue
to work perfectly well on most of the universe.   My objection is not any one you anticipated, 
I didn't expect you to use one of the ones I listed, that's why Isaid "before others jump in"... ;-)Shane

To unsubscribe, change your address, or temporarily deactivate your subscription, 
please go to http://v2.listbox.com/member/[EMAIL PROTECTED]

Re: [agi] Marcus Hutter's lossless compression of human knowledge prize

[Shane Legg]

On 8/13/06, Pei Wang <[EMAIL PROTECTED]> wrote:
> Hutter's only assumption about AIXI is that the environment can be simulated by a Turing machine.That is already too strong to me. Can our environment be simulated bya Turing machine? Why?
There are currently no widely accepted laws of physics that cannotbe simulated by a Turing machine (in the general sense of the word,i.e. I include monotone machines with access to a source of randombits) to an arbitrary degree of accuracy.
Of course this doesn't "prove" that the universe is computable, howeverit is enough evidence to put the onus onto you to prove otherwise.Before some other people jump in with the usual objections (by the
way the perspective I am advocating is the mainstream one) here'ssome replies:Q: What about quantum theory and quantum computers?They are both Turing computable (with great effort I might add).
Quantum computers can do things with lower time complexity,but they can't perform incomputable computations.Q: What about chaos?Again, Turing computable but with great effort.  You might needthe initial conditions to an amazing accuracy and a vast amount
of computation time, but there is no indication that what happensis in any way fundamentally incomputable.Shane

To unsubscribe, change your address, or temporarily deactivate your subscription, 
please go to http://v2.listbox.com/member/[EMAIL PROTECTED]

Re: [agi] Marcus Hutter's lossless compression of human knowledge prize

[Shane Legg]

But Shane, your 19 year old self had a much larger and more diversevolume of data to go on than just the text or speech that you
ingested...I would claim that a blind and deaf person at 19 could pass aTuring test if they had been exposed to enough information overthe years.  Especially if they had the ability to read everything
that was ever spoken to them.  So I don't see why you wouldneed a corpus billions of times larger than this, as you suggested. 
And, of course, your ability to predict your next verbal response isNOT a good indicator of your ability to adaptively deal with newsituations...All I'm talking about is predicting well enough to pass a Turing
test... that was my claim:  That with an amazingly good compressormy life's spoken and written words you could construct a machinethat would pass a Turing test. 
I do not assume that an outstanding compressor of your verbal inputsand outputs would necessarily be a great predictor of your futureverbal inputs and outputs -- because there is much more to you thanverbalizations.  It might make bad errors in predicting your responses
in situations different from ones you had previously experienced... orin situations similar to situations you had previously experienced butthat did not heavily involve verbiage...But if it can't make good predictions to random questions given to
me in a Turing test, then it's not an "amazingly good compressor"of the first 20 years of my life.  Indeed the first 20 years of my lifewould involve tens of thousands of conversations, and I presume on
all of them my responses would have been good enough to pass aTuring test.Shane

To unsubscribe, change your address, or temporarily deactivate your subscription, 
please go to http://v2.listbox.com/member/[EMAIL PROTECTED]

Re: [agi] Marcus Hutter's lossless compression of human knowledge prize

[Shane Legg]

Yes, I think a hybridized AGI and compression algorithm could dobetter than either one on its own  However, this might result in
an incredibly slow compression process, depending on how fast the AGIthinks.  (It would take ME a long time to carry out this process overthe whole Hutter corpus...)Estimate the average compression by sampling.
 Also, not all narrow-AI compression algorithms will necessarily beable to produce output in the style you describe above.  Standard LZ
Sure, use a PPM compressor perhaps. > At least I think you'll agree that an amazingly outstanding Hutter
> test result (possibly on an even larger text corpus that included> conversations etc.) would allow you to then construct a machine> that would pass the Turing test?I agree ONLY in the context of a vastly larger text corpus --- and I
wonder just how large a corpus would be required ... quite possibly,one much larger than all text ever produced in the history of thehuman race...I don't think it's anywhere near that much.  I read at about 2 KB
per minute, and I listen to speech (if written down as plain text)at a roughly similar speed.  If you then work it out, buy the timeI was 20 I'd read/heard not more than 2 or 3 GB of raw text.If you could compress/predict everything that I'd read or heard
until I was 20 years old *amazingly well*, then I'm sure you'dbe able to use this predictive model to easily pass a Turing test.Indeed it's trivially true: Just have me sit a Turing test when Iwas 19.  Naturally I would have passed it, and thus so would
the compressor/predictor (assuming that it's amazingly good,or at least as good at predicting my responses as I would be).Shane

To unsubscribe, change your address, or temporarily deactivate your subscription, 
please go to http://v2.listbox.com/member/[EMAIL PROTECTED]

Re: [agi] Marcus Hutter's lossless compression of human knowledge prize

[Shane Legg]

That seems clear.Human-level AGI =/=> Good Hutter test result

just asHuman =/=> Good Hutter test resultMy suggestion then is to very slightly modify the test as follows:  Instead of just getting the raw characters, what you get is thesequence of characters and the probability distribution over the
next character as predicted by a standard compressor.  You(meaning the algorithm or person being tested) can then chooseto modify this distribution before it is used for compression.So, for example, when the compressor is extremely certain that
the next characters are "do its thing.  But when the string so far is "3x7=" and the compressordoesn't seem to know what the next characters are, you push the
compressor in the right direction.I'm pretty sure that such a combination would easily beat the bestcompressors available when used with a human, or a human levelAGI with world knowledge for that matter.  Indeed I think somebody
has already done something like this before with humans.  Maybeone of the references that Matt gives above.
However, I am uncertain whether
Amazingly outstanding Hutter test result ==> powerful AGIAt least I think you'll agree that an amazingly outstanding Huttertest result (possibly on an even larger text corpus that included
conversations etc.) would allow you to then construct a machinethat would pass the Turing test?Shane

To unsubscribe, change your address, or temporarily deactivate your subscription, 
please go to http://v2.listbox.com/member/[EMAIL PROTECTED]

Re: [agi] Marcus Hutter's lossless compression of human knowledge prize

[Shane Legg]

Ben,So you think that, Powerful AGI  ==> good Hutter test resultBut you have a problem with the reverse implication,good Hutter test result =/=> Powerful AGIIs this correct?
Shane

To unsubscribe, change your address, or temporarily deactivate your subscription, 
please go to http://v2.listbox.com/member/[EMAIL PROTECTED]

Re: [agi] [META] Is there anything we can do to keep junk out of the AGI Forum?

[Shane Legg]

Basically, as you can all probably see, Davy has written a chat bot typeof program.  If you email him he'll send you a copy --- he says it's a bitover 1.5 MB and runs on XP.It's a bit hard to understand how it works, partly because (by his own
confession) he doesn't know much about AI and so doesn't know toproperly describe what he's doing.The machine translation is made considerably worse by the fact that he'snot writing in "proper Italian" --- he's using abbreviations for words, not
using standard vowel accents, punctuation, capitalisation etc...ShaneOn 7/26/06, Davy Bartoloni - Minware S.r.l. <
[EMAIL PROTECTED]> wrote:i Used a very Poor Translator...- Original Message -
From: "BillK" <[EMAIL PROTECTED]>To: Sent: Wednesday, July 26, 2006 4:08 PMSubject: Re: [agi] [META] Is there anything we can do to keep junk out of
the AGI Forum?> On 7/26/06, Richard Loosemore wrote:> >> > I am beginning to wonder if this forum would be better off with a> > restricted membership policy.> >
> >> > Richard Loosemore> >> >> > Davy Bartoloni - Minware S.r.l. wrote:> > > Which thing we want from a IA? , we want TRULY something? thedoubt> > > rises me that nobody affidera' never to the words of a program, e'
piu' easy> > > to entrust itself to a book.   the book from the emergency, its wordscannot> > > change, what there e' written e' nearly sure reliable. it isentrusted> > > to us of piu' to the handbook of the bimby in order making the mousse
of> >>> It *might* sound a bit better in the original Italian???> But Babelfish / Google translate has made a real mess of it.>> BillK>> ---
> To unsubscribe, change your address, or temporarily deactivate yoursubscription,> please go to http://v2.listbox.com/member/[EMAIL PROTECTED]
 --> No virus found in this incoming message.> Checked by AVG Free Edition.> Version: 7.1.394 / Virus Database: 268.10.4/399 - Release Date: 25/07/2006>>
---To unsubscribe, change your address, or temporarily deactivate your subscription,please go to http://v2.listbox.com/member/[EMAIL PROTECTED]


To unsubscribe, change your address, or temporarily deactivate your subscription, 
please go to http://v2.listbox.com/member/[EMAIL PROTECTED]

Re: [agi] Flow charts? Source Code? .. Computing Intelligence? How too? ................. ping

[Shane Legg]

On 7/25/06, Ben Goertzel <[EMAIL PROTECTED]> wrote:
Hmmm...About the measurement of general intelligence in AGI's ...I would tend to advocate a "vectorial intelligence" approachI'm not against a vector approach.  Naturally every intelligent
system will have domains in which it is stronger than others.Knowing what these are is useful and important.  A single numbercan't capture this. 
One might also define a "domain-transcending intelligence", measuredby supplying a system with tasks involving learning how to solveproblems in totally new areas it has never seen before.  This would be
very hard to measure but perhaps not impossible.However, in my view, this "domain-transcending intelligence" -- thoughperhaps the most critical part of "general intelligence" -- should
I think this "most critical part", as you put it, is what's missing ina lot of AI systems.  It's why people look at a machine that cansolve difficult calculus problems in the blink of a second and say
that it's not really intelligent.This is the reason I think there's value in having an overall generalmeasure of intelligence --- to highlight the need to put the G backinto AI.Shane

To unsubscribe, change your address, or temporarily deactivate your subscription, 
please go to http://v2.listbox.com/member/[EMAIL PROTECTED]

Re: [agi] Flow charts? Source Code? .. Computing Intelligence? How too? ................. ping

[Shane Legg]

On 7/24/06, Eugen Leitl <[EMAIL PROTECTED]> wrote:
On Sat, Jul 22, 2006 at 07:48:10PM +0200, Shane Legg wrote:>After some months looking around for tests of intelligence for>machines what I foundWhy would machines need a different test of intelligence than
people or animals?When measuring the intelligence of a human or other animal youhave to use an appropriate test -- clearly cats can't solve linguisticproblems and even if they could they can't use a pen to write down
their answer.  Thus intelligence tests need to take into account theenvironment that the agent needs to deal with, the ways in which itcan interact with its environment, and also what types of cognitiveabilities might reasonably be expected.  However it seems unlikely
that AIs will be restricted to having senses, cognitive abilities orenvironments that are like those of humans or other animals.  Assuch the ways in which we measure intelligence, and indeed ourwhole notion of what intelligence is, needs to be expanded to
accommodate this.Shane

To unsubscribe, change your address, or temporarily deactivate your subscription, 
please go to http://v2.listbox.com/member/[EMAIL PROTECTED]

Re: [agi] Flow charts? Source Code? .. Computing Intelligence? How too? ................. ping

[Shane Legg]

On 7/13/06, James Ratcliff <[EMAIL PROTECTED]> wrote:
On the topic of measuring intelligence, what do you think about the actual structure of comparison of some of today's AI systems.  I would like to see someone come up with and get support for a general fairly widespread set of test s for general AI other than the turing test.
After some months looking around for tests of intelligence for machines what I found was... not very much.  Few people have proposed tests of intelligence for machines,and other than the Turing test, none of these test have been developed or used much.
Naturally I'd like "universal intelligence", that Hutter and myself have formulated,to lead to a practical test that was widely used.  However making the test practicalposes a number of problems, the most significant of which, I think, is the sensitivity
that universal intelligence has to the choice of reference universal Turing machine.Maybe, with more insights, this problem can be, if not solved, at least dealt with ina reasonably acceptable way?Shane


To unsubscribe, change your address, or temporarily deactivate your subscription, 
please go to http://v2.listbox.com/member/[EMAIL PROTECTED]

Re: [agi] Flow charts? Source Code? .. Computing Intelligence? How too? ................. ping

[Shane Legg]

On 7/13/06, Pei Wang <[EMAIL PROTECTED]> wrote:
Shane,Do you mean Warren Smith?Yes.Shane

To unsubscribe, change your address, or temporarily deactivate your subscription, 
please go to http://v2.listbox.com/member/[EMAIL PROTECTED]

Re: [agi] Flow charts? Source Code? .. Computing Intelligence? How too? ................. ping

[Shane Legg]

James,Currently I'm writing a much longer paper (about 40 pages) on intelligencemeasurement.  A draft version of this will be ready in about a month whichI hope to circulate around a bit for comments and criticism.  There is also
another guy who has recently come to my attention who is doing verysimilar stuff.  He has a 50 page paper on formal measures of machineintelligence that should be coming out in coming months.I'll make a post here when either of these papers becomes available.
Shane

To unsubscribe, change your address, or temporarily deactivate your subscription, 
please go to http://v2.listbox.com/member/[EMAIL PROTECTED]

Re: [agi] Flow charts? Source Code? .. Computing Intelligence? How too? ................. ping

[Shane Legg]

On 7/6/06, William Pearson <[EMAIL PROTECTED]> wrote:
> We're good at learning from our environment, but that only gets you> so far, by itself it won't let you do any of the above things because you'll> be dead before you get the hang of them.So this whittles away AIXI and similar formalisms from the possible
candidates for being a pgpps.Even an AIXI needs a "safe" environment to start with, just like a humanbaby.  It would only be after it had learnt the basics of how its environment
works that it would become truly super powerful.When thinking about how powerful an AIXI is I think it is useful to make adistinction between a "baby AIXI", and a "mature AIXI" that had learnt about
its environment.While AIXI is all a bit pie in the sky, "mathematical philosophy" if you like,I think the above does however highlight something of practical importance:Even if your AI is incomputably super powerful, like AIXI, the training and
education of the AI is still really important.  Very few people spend timethinking about how to teach and train a baby AI.  I think this is a greatlyignored aspect of AI.Shane

To unsubscribe, change your address, or temporarily deactivate your subscription, 
please go to http://v2.listbox.com/member/[EMAIL PROTECTED]

Re: [agi] Universal Test for AI?...... AGI bottlenecks

[Shane Legg]

For a universal test of AI, I would of course suggest "universal intelligence"as defined in this report:http://www.idsia.ch/idsiareport/IDSIA-10-06.pdf
ShaneOn Fri, 02 Jun 2006 09:15:26 -500, [EMAIL PROTECTED] <[EMAIL PROTECTED]
> wrote:What is the universal test for the ability of any given AI SYSTEM
to Perceive Reason and Act?Is there such a test?What is the closest test known to date?Dan Goe

To unsubscribe, change your address, or temporarily deactivate your subscription, 
please go to http://v2.listbox.com/member/[EMAIL PROTECTED]

Re: [agi] AGIRI Summit

[Shane Legg]

I note that the human brain seems to store declarative and proceduralknowledge separately and in somewhat different ways --- probably for
reasons related to those underlying our choice in Novamente.

I'm not aware of evidence that the brain stores declarative and procedural
knowledge separately (I'm not a neuroscience expert).  I have however
seen good evidence that formation of declarative and procedural memories
involve different mechanisms.  You can show this in mice by knocking out
parts of the hippocampus and observing how a mouse's learning is affected.

The neuroscientists doing this research believe that a number of learning
mechanisms are at work.  They also show that when you knock one of them
out, the others can compensate, but the learning rate is much much slower.

Shane


To unsubscribe, change your address, or temporarily deactivate your subscription, 
please go to http://v2.listbox.com/member/[EMAIL PROTECTED]

Re: [agi] Who's watching us?

[Shane Legg]

After a few hours digging around on the internet, what I found was thata number of popular blogs get hits from military DNSs.  The most likelyreason seems to be that some people in the military who have office jobs
spend a lot of time surfing the net.  When they find something cool theytell all their other office worker military friends and before you know it,you're getting hits from 10 different military organisations.
I'd never really thought about it before, but I guess people working in thePentagon get bored and read blogs during the day too  ;-)I'll take off my tin foil hat now...Shane



To unsubscribe, change your address, or temporarily deactivate your subscription, 
please go to http://v2.listbox.com/member/[EMAIL PROTECTED]

Re: [agi] Who's watching us?

[Shane Legg]

Daniel,

It seems to be a combination of things.  For example, my most recent
hits from military related computers came from an air force base just
a few hours ago:
px20o.wpafb.af.mil - - [19/Dec/2005:12:07:41 +] "GET /documents/42.pdf HTTP/1.1" 200 50543 "-" "Mozilla/4.0 (compatible; MSIE 6.0; Windows NT 5.1; SV1; .NET CLR 
1.1.4322)"px20o.wpafb.af.mil - - [19/Dec/2005:12:07:58 +] "GET /documents/f92-legg.pdf HTTP/1.1" 200 206778 "-" "Mozilla/4.0 (compatible; MSIE 
6.0; Windows NT 5.1; SV1; .NET CLR 1.1.4322)"px20o.wpafb.af.mil - - [19/Dec/2005:12:08:02 +] "GET /documents/IDSIA-04-04.pdf HTTP/1.1" 200 338053 "-" "Mozilla/4.0 (compatible; MSIE 
6.0; Windows NT 5.1; SV1; .NET CLR 1.1.4322)"px20o.wpafb.af.mil - - [19/Dec/2005:12:08:10 +] "GET /documents/calude99solving.pdf HTTP/1.1" 200 192729 "-" "Mozilla/4.0 (compatible; MSIE 
6.0; Windows NT 5.1; SV1; .NET CLR 1.1.4322)"px20o.wpafb.af.mil - - [19/Dec/2005:12:08:27 +] "GET /documents/96JC-SL-IHW-MDL.pdf HTTP/1.1" 200 101489 "-" "Mozilla/4.0 (compatible; MSIE 
6.0; Windows NT 5.1; SV1; .NET CLR 1.1.4322)"px20o.wpafb.af.mil - - [19/Dec/2005:12:08:31 +] "GET /documents/smith94objective.pdf HTTP/1.1" 200 42124 "-" "Mozilla/4.0 (compatible; MSIE 
6.0; Windows NT 5.1; SV1; .NET CLR 1.1.4322)"px20o.wpafb.af.mil - - [19/Dec/2005:12:08:51 +] "GET /documents/disSol.pdf HTTP/1.1" 200 202651 "-" "Mozilla/4.0 (compatible; MSIE 
6.0; Windows NT 5.1; SV1; .NET CLR 1.1.4322)"

I'm not sure if that's a spider or not.  Some of my other visitors:
 hqda.pentagon.mil lanl.gov defence.gov.au robins.af.mil 
lockheedmartin.com
The Pentagon looked like a spider as they hit multiple pages and
selected out particular images and files all in one second.  Other
patterns of hits look like people with web browsers looking around.

Shane




To unsubscribe, change your address, or temporarily deactivate your subscription, 
please go to http://v2.listbox.com/member/[EMAIL PROTECTED]

Re: [agi] Who's watching us?

[Shane Legg]

Jiri,

I would have assumed that to be the case, like what Ben said.

I guess they have just decided that my research is sufficiently
interesting to keep up to date on.  Though getting hits from these
people on a daily basis seems a bit over the top.  I only publish
something once every few months or so!

Shane




To unsubscribe, change your address, or temporarily deactivate your subscription, 
please go to http://v2.listbox.com/member/[EMAIL PROTECTED]

Re: [agi] neural network and AGI

[Shane Legg]

Pei,> What you seem to be criticising in your memo is what I'd call> "feed forward neural networks".
I see what you mean, though in the memo I didn't rule out feedback.Recurrence makes all the difference...For example, consider a very simple neural network model:Rational activations, fixed topologies, no learning rules or weight
changes allowed, and just a trivial linear activation function.If you allow only feed forward connections then such a networkcan only compute linear functions of inputs.  Pretty dumb.On the other hand,  if you allow these networks to have recurrent
connections, the model is in fact Turing complete.  Indeed peoplehave built networks in this model that simulate classical universalTuring machines (see the work of Siegelman).  There are evencompilers for high level languages like Occam that will output
a recurrent neural network to execute your program (see the workof Neto for example).Anyway my point is, once you allow recurrent connections eventrivial types of neural networks can, in theory, compute anything.
This is why I prefer to think of NNs as a computational paradigmrather than a class of techniques, algorithms or methods.You could argue that NARS is a better way of thinking about orexpressing AGI, or something like that, perhaps.  Just as you
might argue that C is a better way of programming than machinecode.  However the limitations aren't fundamental, indeed I couldwrite a NARS system in Occam and then compile it to run asa neural network.
Shane



To unsubscribe, change your address, or temporarily deactivate your subscription, 
please go to http://v2.listbox.com/member/[EMAIL PROTECTED]

Re: [agi] neural network and AGI

[Shane Legg]

Pei,To my mind the key thing with neural networks is that theyare based on large numbers of relatively simple units thatinteract in a local way by sending fairly simple messages.Of course that's still very broad.  A CA could be considered
a neural network according to this description, and indeedto my mind I don't see much difference between the two.Nevertheless, it does rule out many things --- just talk tosomebody who has tried to take a normal algorithm that
does something and turn it into an algorithm that workson a massively parallel architecture using relatively simplecomputations units.As I see it neural networks are more a paradigm of computationrather than any specific AI method or technology.  This means
that talking about them in general is difficult.What you seem to be criticising in your memo is what I'd call"feed forward neural networks".Shane



To unsubscribe, change your address, or temporarily deactivate your subscription, 
please go to http://v2.listbox.com/member/[EMAIL PROTECTED]

[agi] Who's watching us?

[Shane Legg]

My AGI ideas continue to brew away, as they have done for many years. I like to keep a fairly low profile for a few reasons, one being that every yearI realise how little I knew the previous year!  I am however busy in my thesis
area of theoretical models of universal super intelligence (AIXI etc.) and sopublish and turn up in magazines etc. from time to time.  I also write a blogthat gets a fair amount of daily traffic.Recently, however, it seems that I've somehow come to the attention of
certain groups.  To start with I thought that the hits I was getting from variousmilitary and military related companies were just random and part of thegrowing number of hits that my blog and website were getting.  But things
have since become more systematic, with military related aerospacecompanies, various military bases and even a certain five sided buildingdoing automated scans of my research site, downloading my papers and
also scanning my blog.Yeah I know, that sounds crazy.  I mean, if these people wanted to watchme, they wouldn't be scanning my stuff using a computer with a Pentagonor other high profile DNS addresses right?  I'm sure that if they really wanted
to watch me they would be far more sneaky and I'd never know about it. I can only presume that for what ever reason, they don't mind me knowingthat I'm being watched.  I've also found out through other channels that I
have attracted some interest.  So they aren't being very secret about it.Anyhow, I presume that the reason I'm attracting attention has somethingto do with my research interests in 
models of super intelligence, AGI andso on.  I can't imagine that anything else about me could be of the slightestinterest to them at all.

Which brings me to my question:  Are we (people interested in AGI)being watched?I'd always just presumed that we weren't considered very interesting,at least, not yet.Shane



To unsubscribe, change your address, or temporarily deactivate your subscription, 
please go to http://v2.listbox.com/member/[EMAIL PROTECTED]

Re: [agi] neural network and AGI

[Shane Legg]

Hi Pei,Most of our disagreement seems to be about definitions and choicesof words, rather than facts.

(1) My memo is not intend to cover every system labeled as "neural network"--- that is why I use a whole section to define what I mean by "NNmodel" discussed in the paper. I'm fully aware of the fact that given
...My strategy is to first discuss the most typical models of the "neural
network" family (or the "standard NN architectures", as Ben put it),My problem is this: At my research institute a large portion of the people workon neural networks.  Things like, recurrent LSTM networks for continuous
speech recognition and evolved echo state networks for real time adaptivecontrol problems.  I also follow research on computational liquids, biologicallyplausible neural networks, neural microcircuit research, and the ideas of
people like Jeff Hawkins.  In my mind, this is all work on "neural networks", andthe researchers themselves call it that, and publish in big NN conferences likeICANN, IJCNN and journals like "Neural Networks".  However none of this work
is like the "NN model" you have defined.  Thus to my mind, your "NN model"does not represent modern neural network research. 
(3) Neuroscience results cannot be directly used to support"artificial neural networks"I disagree as a number of the trends and new ideas in artificial neural networksthat I follow are coming from neuroscience research.
If I had to sum up our differences: I'd say that what you call "standard neuralnetworks" and your "NN model", and most of the problems you describe, wouldhave been reasonable in 2000... but not now,  5 to 6 years later.
Shane



To unsubscribe, change your address, or temporarily deactivate your subscription, 
please go to http://v2.listbox.com/member/[EMAIL PROTECTED]

Re: [agi] neural network and AGI

[Shane Legg]

Ben,My suspicion is that in the brain knowledge is often stored on two levels:
* specific neuronal groups correlated with specific informationIn terms of the activation of specific neurons indicating high level concepts,I think there is good evidence of this now.  See for example the work of
Gabriel Kreiman.* strange attractors spanning large parts of the brain, correlated
with specific informationThe liquid computation model of Maass, Natschlager and Markram, which wasinspired by neuroscience work on neural microcircuits and cortical columns,shows how  complex dynamics can be used to perform useful computations
from random networks.  Jaeger has studied how, essentially the same model,can be used as a short term memory using attractors in the system.  Thus itseems reasonable that, at least at the columnar scale, attractors could play
an important role in short term memory.Shane



To unsubscribe, change your address, or temporarily deactivate your subscription, 
please go to http://v2.listbox.com/member/[EMAIL PROTECTED]

Re: [agi] neural network and AGI

[Shane Legg]

Hi Pei,As usual, I disagree! I think you are making a straw man argument.The problem is that what you describe as neural networks is just a certainlimited class of neural networks.  That class has certain limitations, which
you point out.  However you can't then extend those conclusions to neuralnetworks in general.  For example...You say, "Starting from an initial state determined by an input vector..."For recurrent NNs this isn't true, or at least I think that your description
is confusing.  The state is a product of the history of inputs, rather thanbeing determined by "an input vector".  Similarly I also wouldn't say thatNNs are about input-output function learning.  Back prop NNs are about
this when used in simple configurations.   However this isn't true of NNsin general, in particular its not true of recurrent NNs.  See for exampleliquid machines or echo state networks.I also wouldn't be so sure about neurons not being easily mapped to
conceptual units.  In recent years neuro scientists have found thatsmall groups of neurons in parts of the human brain correspond to veryspecific things.  One famous case is the "Bill Clinton neuron".  Of course
you're talking about artificial NNs not real brains.  Nevertheless, if biologicialNNs can have this quasi-symbolic nature in places, then I can't see howyou could argue that artificial NNs can't do it due to some fundamental
limitation.I have other things to say as well, but my main problem with the paperis what I've described above.  I don't think your criticisms apply to neuralnetworks in general.Shane



To unsubscribe, change your address, or temporarily deactivate your subscription, 
please go to http://v2.listbox.com/member/[EMAIL PROTECTED]

[agi] Open AGI?

[Shane Legg]

Hi all,

I'm curious about the general sentiments that people have
about the appropriate level of openness for an AGI project.
My mind certainly isn't made up on the issue and I can see
reasons for going either way.  If a single individual or
small group of people made a sudden break through in AGI
design this would place a huge amount of power in their
hands.  I could easily see this situation being dangerous.
On the other hand I'm not sure that I'd want too many people
knowing how to do it either!  Already the world seems to
have a few too many people who have the detailed knowledge
required to build a working nuclear weapon for example.
What are your thoughts?  Surely this has been debated
many times before I suppose?
Cheers
Shane
---
To unsubscribe, change your address, or temporarily deactivate your subscription, 
please go to http://v2.listbox.com/member/[EMAIL PROTECTED]

Re: [agi] Complexity of Evolving an AGI

[Shane Legg]

Hi Ben,

Thanks for the comments.

I understand your perspective and I think it's a reasonable one.
Well your thinking has surely left it's mark on my views ;-)


I think that what you'll get from this approach, if you're lucky, is a kind
of "primitive brain", suitable to control something with general
intelligence around that of a reptile or a very stupid mammal.
Then, you can use the structures/dynamics of this primitive brain as raw
materials, for constructing a more powerful general intelligence.
Yes, this is the basic idea.  Some structures and dynamics will tend
to categorize and cluster information, some will tend to recognize
temporal patterns, some might tend to store information, some will
be mixtures of these things.  Exactly what I'll find is hard to know
in advance.  As you say this isn't a powerful "brain", but it might
serve as a set of useful building blocks to construct more powerful
general intelligence from.
Your approach seem to be to skip this first step by careful design.
The danger is that it's hard and you might mess it up.  Some of the
needed dynamics could be very subtle and easily missed.  Also it's
not clear just how "small" the fundamental units of representation
have to be in order to be flexible and dynamic enough.
The danger with my approach is that my theory might be too weak.
This could leave me with a search space that is too big to effectively
search.  Also, if it turns out that my fundamental units of
representation are smaller than what is really needed, I might produce
very inefficient solutions.  I also need to look more closely at
why all the researchers who were evolving neural network dynamics
seem to have given up and are now doing other things... that's not
a good sign!  One of the main guys who was doing this is at a research
institute near here so I might have to go visit him for a chat about it.
Heh, our approaches slowly get closer.  Novamente is greatly simplified
and cleaned up from the 1.5 million lines of code that was webmind, while
vetta has very slowly been increasing in complexity since my webmind days.
Still, this is only a weekend hobby for me --- I need to focus my energies
on getting my PhD done.  Building an AGI might have to wait.
ciao
Shane
---
To unsubscribe, change your address, or temporarily deactivate your subscription, 
please go to http://v2.listbox.com/member/[EMAIL PROTECTED]

Re: [agi] Complexity of Evolving an AGI

[Shane Legg]

Ciao,

Is something weird going on with the AGI list here?  I just got two
emails claiming to be from a month or so ago that where actually
sent today...
Anyway in reply to Ben's email,

Ben Goertzel wrote:
But the different trials need not be independent --- we can save the
trajectory of each AI's development continuously, and then restart a new
branch of "AI x at time y" for any recorded AI x at any recorded time point
y.
Also, we can intentionally form composite AI's by taking portions of AI x's
mind and portions of AI y's mind and fusing them together into a new AI z...
So we don't need to follow a strict process of evolutionary trial and error,
which may accelerate things considerably  particularly if, as
experimentation progresses, we are able to learn abstract theories about
what makes some AI's smarter or stabler or friendlier than others.
I agree totally.  Indeed I advocate going further and actually evolving
the fundamental structures and dynamics that drive the system --- designing
them by hand or trying to prove any useful results about what happens in
a complex recurrent network seems to be really difficult.  Thus perhaps a
combination of artificial evolution, experimentation, and the development
of theories to explain what we see is the most likely approach to succeed.
At least that's my best guess at the moment based on what I've seen working
on various AI/AGI projects in the past.
I sent Ben an email along similar lines a few days back describing my
own little (extremely slowly moving and incomplete) set of AGI ideas that
I refer to as the vetta project.  I've pasted part of what I wrote below
for anybody who is interested.
Cheers
Shane
---

Well it's a mix of things really --- and it changes over time a bit too!

Basically my approach goes something like this:

1) Build a set of precise "IQ" tests for machines.  These tests cover
   everything from bacteria level intelligence to super human intelligence.
   It's a reasonably complex web of relations: passive predictors, classifiers,
   simple reactive systems, Markov chains, MDPs, POMDPs and many others.
   You can prove a whole bunch of relations between all these mathematically,
   indeed that's what I did for the first 4 months of my PhD.  That's the first
   step; however it doesn't really capture how difficult a problem is.  So for
   that you need something like like complexity theory (both time and space).
   Anyway, the point is that you can then measure exactly where in this complex
   tree of abilities an AGI system is.  The most general form of this is what
   I call "cybernance" and is closely related to the "intelligence order
   relation" that appears in the AIXI proofs.
2) Define a space of systems that should contain an AGI.  This is a bit harder
   to explain.  Again complexity theory comes into it.  So things like the
   fact that I think that the "meta logic" of a system has to be very small
   and thus the building blocks of the system must be quite simple.  Also that
   the processing of the system must have certain self-organizing properties
   such as compression of information in space and time, consistency over levels
   of abstraction and stuff like that.  This is the more philosophical part I
   suppose.  The point is that I need to make this space of possible systems
   as small as I can without making a mistake and excluding a working design
   for an AGI from the set.  Oh, and I should mention that I'm thinking of
   some kind of information processing network here: some kind of neural network,
   Hebbian network, HMM, Bayesian network.  Basically the space is a super set
   of all these things and more.
3) Genetic programming.  (1) gives us a fine grained multi-objective fitness
   function and (2) defines a search space.  Now I can't just run my GA and
   expect things to work here!  Clearly the space in (2) is going to be pretty
   large.  So at this point it becomes a bit of an experimental science and I
   have to mix things around a bit.  So I'll be restricting the tests to just
   certain very simple objectives and restricting the space to smaller subspaces
   to see what works and what doesn't.  Then try to cross over solutions to find
   systems that work for both etc.  Hopefully at this stage I can zero in on
   promising parts of the space of possible designs.  Perhaps even design my
   own attempts at functioning systems and throw them into the evolutionary mix
   and see if they can breed with other different partial solutions to form
   new and interesting things.
I guess in a sense it's the natural evolution of intelligence but on steroids:
rather than having fitness related to intelligence very indirectly via survival
here we measure a kind of computational intelligence very direction and equate it
with survival.  Also we restrict the space of possible designs as much as we can
get away with to speed things up --- this 

Re: [agi] FYI: AAAI Symposium on Human-Level Intelligence

[Shane Legg]

Also I think this is pretty cool in case you miss it:

http://www.ai.mit.edu/projects/genesis/movies.html

---
To unsubscribe, change your address, or temporarily deactivate your subscription, 
please go to http://v2.listbox.com/member/[EMAIL PROTECTED]

Re: [agi] FYI: AAAI Symposium on Human-Level Intelligence

[Shane Legg]

Thanks Pei.

Following the links to the people who are running this I found a
whole bunch of academic AI people who are interested in and working
on general intelligence.  Their approach seems to be very much
based around the idea that powerful systems for vision, sound, speech,
motor skills and so on form the basis for general intelligence.
Somehow I'd managed to miss seeing these people and their projects
in the past.  For example:
http://www.ai.mit.edu/projects/genesis/

http://www.ai.mit.edu/projects/HIE/index.html

http://www.cassimatis.com/

Shane

Pei Wang wrote:

2004 AAAI Fall Symposium Series
Achieving Human-Level Intelligence through Integrated Systems and Research
October 21-24, 2004
Washington D.C.
See http://xenia.media.mit.edu/~nlc/conferences/fss04.html

---
To unsubscribe, change your address, or temporarily deactivate your subscription, 
please go to http://v2.listbox.com/member/[EMAIL PROTECTED]

---
To unsubscribe, change your address, or temporarily deactivate your subscription, 
please go to http://v2.listbox.com/member/[EMAIL PROTECTED]

[agi] Two nice non-technical articles

[shane legg]

Agi types might like these two articles,

http://www.theregister.co.uk/content/4/33463.html

http://www.theregister.co.uk/content/4/33486.html

Shane



Want to chat instantly with your online friends?  Get the FREE Yahoo!
Messenger http://mail.messenger.yahoo.co.uk

---
To unsubscribe, change your address, or temporarily deactivate your subscription, 
please go to http://v2.listbox.com/member/[EMAIL PROTECTED]

Re: [agi] Complexity of environment of agi agent

[Shane Legg]

Ciao Arnoud,

Perhaps my pattern wasn't clear enough

1
2
3
4
.
.
.
00099
00100
00101
.
.
.
0
1
.
.
.
8
9
then repeat from the start again.  However each character is
part of the sequence.  So the agent sees 10002300...
So the "whole pattern" in some sense is 100,000 "numbers"
each of 5 characters giving a 500,000 character pattern
of digits from 0 to 9.  A human can learn this reasonably
easily but your AI won't.  It would take something more
like a mega byte to store the pattern.  Actually with
the overhead of all the rules it would be much bigger.
Shane

---
To unsubscribe, change your address, or temporarily deactivate your subscription, 
please go to http://v2.listbox.com/member/[EMAIL PROTECTED]

Re: [agi] Complexity of environment of agi agent

[Shane Legg]

Arnoud,

I'm not sure if this makes much sense.  An "ideal" agent is not going
to be a "realistic" agent.  The bigger your computer and the better
your software more complexity your agent will be able to deal with.
With an ideal realistic agent I meant the best software we can make on the 
best hardware we can make.
In which case I think the question is pretty much impossible to
answer.  Who knows what the best hardware we can make is?  Who
knows what the best software we can make is?

Do I have to see it like something that the value of the nth bit is a 
(complex) function of all the former bits? Then it makes sense to me. After 
some length l of the pattern computation becomes unfeasible.
But this is not the way I intend my system to handle patterns. It learns the 
pattern after a lot of repeted occurences of it (in perception). And then it 
just stores the whole pattern ;-) No compression there. But since the 
environment is made outof smaller patterns, the pattern can be formulated in 
those smaller patterns, and thus save memory space.
This is ok, but it does limit the sorts of things that your system
is able to do.  I actually suspect that humans do a lot of very
simple pattern matching like you suggest and in some sense "fake"
being able to work out complex looking patterns.  It's just that
we have seen so many patterns in the past and that we are very
good at doing fast and sometimes slightly abstract pattern matching
on a huge database of experience.   Nevertheless you need to be a
little careful because some very "simple" patterns that don't
repeat in a very explicit way could totally confuse your system:
123.9123

Your system, if I understand correctly, would not see the pattern
until it had seen the whole cycle several times.  Something like
5*100,000*2 = 1,000,000 characters into the sequence and even then
it would need to remember 100,000 characters of information.  A
human would see the pattern after just a few characters with
perhaps some uncertainly as to what will happen after the 9.
The total storage required for the pattern with a human would be
far less than 100,000 characters your system would need too.

Yes, but in general you don't know the complexity of the simplest solution of 
the problem in advance. It's more likely that you get to know first what the 
complexity of the environment is.
In general an agent doesn't know the complexity of its
environment either.

The strategy I'm proposing is: ignore everything that is too complex. Just 
forget about it and hope you can, otherwise it's just bad luck. Of course you 
want to do the very best to solve the problem, and that entails that some 
complex phenomenon that can be handled must not be ignored a priori; it must 
only be ignored if there is evidence that understanding that phenomenon does 
not help solving your the problem.
In order for this strategy to work you need to know what the maximum 
complexity is an agent can handle, as a function of the resources of the 
agent: Cmax(R). And it would be very helpful for making design decisions to 
know Cmax(R) in advance. You can then build in that everything above Cmax(R) 
should be ignored; 'vette pech' as we say in Dutch if you then are not able 
to solve the problem. 
Why not just do this dynamically?  Try to look at how much of
the agent's resources are being used for something and how much
benefit the agent is getting from this.  If something else comes
along that seems to have a better ratio of benefit to resource
usage then throw away some of the older stuff to free up resources
for this new thing.
Shane

---
To unsubscribe, change your address, or temporarily deactivate your subscription, 
please go to http://v2.listbox.com/member/[EMAIL PROTECTED]

Re: [agi] Complexity of environment of agi agent

[Shane Legg]

arnoud wrote:

How large can those constants be? How complex may the environment be maximally 
for an ideal, but still realistic, agi agent (thus not a solomonof or AIXI 
agent) to be still succesful? Does somebody know how to calculate (and 
formalise) this?
I'm not sure if this makes much sense.  An "ideal" agent is not going
to be a "realistic" agent.  The bigger your computer and the better
your software more complexity your agent will be able to deal with.
The only way I could see that it would make sense would be if you
could come up with an algorithm and prove that it made the best
possible usage of time and space in terms of achieving its goals.
Then the constants you are talking about would be set by this
algorithm and the size of the biggest computer you could get.

Not even an educated guess?

But I think some things can be said:
Suppose perception of the environment is just a bit at a time: 
...010100010010010111010101010...

In the random case: for any sequence of length l the number of possible 
patterns is 2^l. Completely hopeless, unless prediction precision need 
decreases also exponentially with l. But that is not realistic. You then know 
nothing, but you want nothing also.
Yes, this defines the limiting case for Solomonoff Induction...


in the logarithmic case: the number of possible patterns of length l increases 
logarithmically with l: #p < constant * log(l). If the constant is not to 
high this environment can be learned easily. There is no need for vagueness 
Not true.  Just because the sequence is very compressible in a
Kolmogorov sense doesn't imply that it's easy to learn.  For example
you could have some sequence where the computation time of the n-th
bit take n^1000 computation cycles.  There is only one pattern and
it's highly compressible as it has a pretty short algorithm however
there is no way you'll ever learn what the pattern is.

I suppose the point I'm trying to make is that complexity of the environment 
is not all. It's is also important to know how many of the complexity can be 
ignored.
Yes.  The real measure of how difficult an environment is is not
the complexity of the environment, but rather the complexity of
the simplest solution to the problem that you need to solve in
that environment.
Shane

P.S. one of these days I'm going to get around to replying to your
other emails to me!!  sorry about the delay!
---
To unsubscribe, change your address, or temporarily deactivate your subscription, 
please go to http://v2.listbox.com/member/[EMAIL PROTECTED]

RE: [agi] Discovering the Capacity of Human Memory

[shane legg]

Yeah, it's a bit of a worry.

By the way, if anybody is trying to look it up, I spelt the guy's
name wrong, it's actually "Stirling's equation".  You can find
it in an online book here:

http://www.inference.phy.cam.ac.uk/mackay/itprnn/book.html

It's a great book, about 640 pages long.  The result I
used is equation 1.13 which is on page 2.

Shane



 --- Brad Wyble <[EMAIL PROTECTED]> wrote: > 
> 
> It's also disconcerting that something like this can make it through the 
> review process.
> 
> Transdisciplinary is oftentimes a pseudonym for combining half-baked and 
> ill-formed ideas from multiple domains into an incoherent mess. 
> 
> This paper is an excellent example.  (bad math + bad neuroscience != good 
> paper)
> 
> 
> 
> 
> 
> ---
> To unsubscribe, change your address, or temporarily deactivate your subscription, 
> please go to http://v2.listbox.com/member/[EMAIL PROTECTED] 


Want to chat instantly with your online friends?  Get the FREE Yahoo!
Messenger http://mail.messenger.yahoo.co.uk

---
To unsubscribe, change your address, or temporarily deactivate your subscription, 
please go to http://v2.listbox.com/member/[EMAIL PROTECTED]

RE: [agi] Discovering the Capacity of Human Memory

[shane legg]

 --- "Amara D. Angelica" <[EMAIL PROTECTED]> wrote: > 
> > 1) The quote to totally wrong the the "^" should be a "," ?
> 
> It's 10 to the 8432 power, according to the paper. This is the
> theoretical memory capacity, not its actual size, but no estimates are
> given for real-world typical size of memory, so the comparison with
> machine capacity seems unrealistic.

Really, they should have realised that 10^8432 bits makes no sense
at all.


> The number is derived as follows:
> 
> "Assuming there are n neurons in the brain, and on average
> there are m connections between a given neuron and the rest of them, the
> magnitude of the brain memory capacity can be expressed by the following
> mathematical model, the human memory capacity model, as given below:
> 
> n!/[m!(n-m)!]
> 
> where n is the total number of neurons and m the number of average
> partial connections between neurons.
> 
> However, this is "extremely hard to calculate and is almost intractable
> using a modern computer, because of the exponential complicity or the
> recursive computational costs for such large n and m," so they did some
> math tricks to estimate it. 

Well the standard way to estimate (sorry about reverting to LaTeX here
for those of you who aren't mathematicians... ) this combination is
to use the log from of Sterling's equation:

\log_2 \binom{n}{m} = (n-m) \log_2 \frac{n}{n-m} + m \log_2 \frac{n}{m}

here n = 10^9 (neurons in brain)
and  m = 10^4 (10,000 connections per neuron)

These are my numbers, if they use slightly different ones it won't
matter too much.

This gives approx,

10^9 \log_2 10^5 + 10^4 \log_2 10^5

which is about 10^10 bits which is about what I'd expect very roughly.
They should have talked to a mathematician first.  Estimating basic
combinatorics like n!/[m!(n-m)!] is not hard.

Shane



Want to chat instantly with your online friends?  Get the FREE Yahoo!
Messenger http://mail.messenger.yahoo.co.uk

---
To unsubscribe, change your address, or temporarily deactivate your subscription, 
please go to http://v2.listbox.com/member/[EMAIL PROTECTED]

Re: [agi] Discovering the Capacity of Human Memory

[shane legg]

Thanks for the link Pei.

The thing is that they are talking about the number of BITS not the
number of POSSIBLE STATES.  Given x bits the number of possible
states is 2^x.  For example with 32 bits you can have 2^32 different
states... or about 4,000,000,000 possible states.  

Thus, if the brain has 10^8432 bits of storage as they claim, then
the number of possible states is 2^(10^8432). 

To make things even worse, even if they realise their error and
decided that they didn't understand what a bit is and that they
actually meant "possible states", the number of bits in this case
then becomes just log_2 (10^8432) = 8432 * log_2 (10) = 28,010 bits
or about 3.5 kilo bytes of storage.  I'd like to think that I have
more than a 3.5 Kb brain!!

They really should have "sanity checked" their results.

Shane


 --- Pei Wang <[EMAIL PROTECTED]> wrote: > The paper can be accessed at
> http://www.enel.ucalgary.ca/People/wangyx/Publications/Papers/B&M-Vol4.2-HMC.pdf
> 
> Their conclusion is based on the assumptions that there are 10^11 neurons
> and their average synapses number is 10^3. Therefore the total potential
> relational combinations is
> (10^11)! / (10^3)! ((10^11)! - (10^3)!), which is approximately 10^8432.
> 
> The model is obviously an oversimplification, and the number is way too big.
> 
> Pei
> 
> - Original Message - 
> From: "shane legg" <[EMAIL PROTECTED]>
> To: <[EMAIL PROTECTED]>
> Sent: Tuesday, September 16, 2003 6:24 AM
> Subject: RE: [agi] Discovering the Capacity of Human Memory
> 
> 
> >
> > The total number of particles in the whole universe is usually
> > estimated to be around 10^80.  These guys claim that the storage
> > of the brain is 10^8432 bits.  That means that my brain has around
> > 10^8352 bits of storage for every particle in the whole universe.
> >
> > I thought I was feeling smarter than usual this morning!
> >
> > Possible explanations:
> >
> > 1) The quote to totally wrong the the "^" should be a "," ?
> >
> > 2) They got confused and thought it was 1 April
> >
> > 3) They are actually doing research into just how flaky AI
> >researchers really are and how easy it is to publish
> >mathematical nonsense in "Mind and Brain" Journal
> >
> > 4) The "scientists" somehow managed to get their PhDs without
> >understanding how numbers work
> >
> > 5) They concluded that the brain is really analogue and so they
> >worked out the volume of the skull at the Planck scale (actually
> >that doesn't work either as the Planck length is far far far to
> >large at 1.6 x 10^-35 m)
> >
> > and so on...
> >
> > Does anybody have a better explanation?
> >
> > Shane
> >
> >
> > --- "Amara D. Angelica" <[EMAIL PROTECTED]> wrote: >
> > http://www.kurzweilai.net/news/news_printable.html?id=2417
> > >
> > > Discovering the Capacity of Human Memory
> > >
> > > Brain and Mind, August 2003
> > >
> > >
> > > The memory capacity of the human brain is on the order of 10^8432 bits,
> > > three scientists have estimated.
> > >
> > > Writing in the August issue of Brain and Mind, their "OAR" cognitive
> > > model asserts that human memory and knowledge are represented by a
> > > network of relations, i.e., connections of synapses between neurons,
> > > rather than by the neurons themselves as in the traditional
> > > information-container model (1 neuron = 1 bit).
> > >
> > > This explains why "the magnitude of neurons in an adult brain seems
> > > stable; however, huge amount of information can be remembered throughout
> > > the entire life of a person," they point out.
> > >
> > > Based on the projected computer memory capacity of 8 x 10^12 bits in the
> > > next ten years, Yingxu Wang et al. conclude that the memory capacity of
> > > a human brain is equivalent to at least "10^8419 modern
> > > computersThis tremendous difference of memory magnitudes between
> > > human beings and computers demonstrates the efficiency of information
> > > representation, storage, and processing in the human brains."
> > >
> > > They also conclude that "this new factor has revealed the tremendous
> > > quantitative gap between the natural and machine intelligence" and that
> > > "next-generation computer memory systems may be built according to their
> > > relational model rath

RE: [agi] Discovering the Capacity of Human Memory

[shane legg]

 
The total number of particles in the whole universe is usually
estimated to be around 10^80.  These guys claim that the storage
of the brain is 10^8432 bits.  That means that my brain has around
10^8352 bits of storage for every particle in the whole universe.

I thought I was feeling smarter than usual this morning!

Possible explanations:

1) The quote to totally wrong the the "^" should be a "," ?

2) They got confused and thought it was 1 April

3) They are actually doing research into just how flaky AI
   researchers really are and how easy it is to publish
   mathematical nonsense in "Mind and Brain" Journal

4) The "scientists" somehow managed to get their PhDs without
   understanding how numbers work

5) They concluded that the brain is really analogue and so they
   worked out the volume of the skull at the Planck scale (actually
   that doesn't work either as the Planck length is far far far to
   large at 1.6 x 10^-35 m)

and so on...

Does anybody have a better explanation?

Shane


--- "Amara D. Angelica" <[EMAIL PROTECTED]> wrote: >
http://www.kurzweilai.net/news/news_printable.html?id=2417  
> 
> Discovering the Capacity of Human Memory
> 
> Brain and Mind, August 2003
> 
> 
> The memory capacity of the human brain is on the order of 10^8432 bits,
> three scientists have estimated. 
> 
> Writing in the August issue of Brain and Mind, their "OAR" cognitive
> model asserts that human memory and knowledge are represented by a
> network of relations, i.e., connections of synapses between neurons,
> rather than by the neurons themselves as in the traditional
> information-container model (1 neuron = 1 bit). 
> 
> This explains why "the magnitude of neurons in an adult brain seems
> stable; however, huge amount of information can be remembered throughout
> the entire life of a person," they point out. 
> 
> Based on the projected computer memory capacity of 8 x 10^12 bits in the
> next ten years, Yingxu Wang et al. conclude that the memory capacity of
> a human brain is equivalent to at least "10^8419 modern
> computersThis tremendous difference of memory magnitudes between
> human beings and computers demonstrates the efficiency of information
> representation, storage, and processing in the human brains." 
> 
> They also conclude that "this new factor has revealed the tremendous
> quantitative gap between the natural and machine intelligence" and that
> "next-generation computer memory systems may be built according to their
> relational model rather than the traditional container metaphor" because
> "the former is more powerful, flexible, and efficient, and is capable of
> generating a mathematically unlimited memory capacity by using limited
> number of neurons in the brain or hardware cells in the next generation
> computers." 
> 
> Brain and Mind 4 (2): 189-198, August 2003
> 
> ---
> To unsubscribe, change your address, or temporarily deactivate your subscription, 
> please go to http://v2.listbox.com/member/[EMAIL PROTECTED] 


Want to chat instantly with your online friends?  Get the FREE Yahoo!
Messenger http://mail.messenger.yahoo.co.uk

---
To unsubscribe, change your address, or temporarily deactivate your subscription, 
please go to http://v2.listbox.com/member/[EMAIL PROTECTED]

Re: [agi] Robert Hecht-Nielsen's stuff

[Shane Legg]

Brad Wyble wrote:
Well the short gist of this guy's spiel is that Lenat is on
> the right track.

My understanding was that he argues that Lenat is on the wrong
track!  Lenat is trying to accumulate a large body of relatively
high level logical rules about the world.  This is very hard to
do and requires a vast amount of human labour to do.
Hecht-Nielsen argues that the brain doesn't in fact logically reason
(at least not directly or course) but rather does something simpler.
Essentially some kind of pattern matching and association as I
understand it.  Also, the knowledge representation system is at a
sub-logical being made up of relatively simple associations formed
on the most part by experience.
This is the complete opposite of Cyc.  Indeed, he attributes much
of the failure of AI to be due to the assumption that intelligent
systems should work on some kind of logic and use logical expressions
to represent knowledge.  Surely Cyc is in this tradition and so,
from Hecht-Nielsen's point of view, must be on the wrong track?
Shane

---
To unsubscribe, change your address, or temporarily deactivate your subscription, 
please go to http://v2.listbox.com/member/[EMAIL PROTECTED]

[agi] Robert Hecht-Nielsen's stuff

[Shane Legg]

A while back Rob Sperry posted a link to a video
of a presentation by Robert Hecht-Nielsen.
( http://inc2.ucsd.edu/inc_videos/ )
In it he claims to have worked out how the brain
thinks :)  I didn't look at it at the time as it's
150MB+ and only had a dial up account, but checked
it out the other day with my new ADSL account.
Aside from his, well, rather over the top style of
presentation, what do people think of this?  I haven't
seen any comment to this list and don't know much
neuroscience myself either.  He has a book too, has
anybody read it?
Cheers
Shane
---
To unsubscribe, change your address, or temporarily deactivate your subscription, 
please go to http://v2.listbox.com/member/[EMAIL PROTECTED]

Re: [agi] Request for invention of new word

[Shane Legg]

Semiborg?

:)

Shane

Ben Goertzel wrote:
Hi ,

For a speculative futuristic article I'm writing (for a journal issue edited
by Francis Heylighen), I need a new word: a word to denote a mind that is
halfway between an individual mind and a society of minds.
Not a hive-mind, but rather a community of minds that exchange
thoughts/ideas directly rather than thru language, and hence derive a degree
of synergetic mutual intelligence much greater than that achievable in a
society of separate minds
I'm reviewing two possible future examples of such minds

1) a community of Novamente AI engines

2) a community of human minds that are enhanced with neurochips or related
technologies and linked into a Net enabled with "Kazaa for
thought-complex-swapping"
Any suggestions?

Also, any reference to prior ("serious" or science-fictional) developments
of this theme?
Thanks,
Ben G
---
To unsubscribe, change your address, or temporarily deactivate your subscription, 
please go to http://v2.listbox.com/member/[EMAIL PROTECTED]



---
To unsubscribe, change your address, or temporarily deactivate your subscription, 
please go to http://v2.listbox.com/member/[EMAIL PROTECTED]

Re: [agi] Hard Wired Switch

[Shane Legg]

To me this approach sounds like you are trying to out smart the AGI;
to somehow constrain it in some way so that it is forced into being
nice to people... or Ben at least.  I don't think this kind of an
approach will work due to the simple fact that you'll never be able
to outsmart a super intelligence.
The trick will be, I think, to dig right down the the deepest essence
of the AGI and make sure that it would never want to harm humanity in
the first place.  I think that's really the only place that danger
could be stopped.  Trying to control the AGI at higher levels seems
doomed to fail to me as the AGI will simply be far to smart to be
restrained by any system that a human could devise.
Exactly what this means is of course hard to establish in the absence
of an obviously correct AGI design... but it's probably worth thinking
a bit about the problem sooner rather than later nevertheless.
I think part of the solution will be to control, at least early on
while the AGI is learning, its interaction with humans.  I read Mary
Shelley's Frankenstein recently and thought that it is still quite
relevant in this respect.  Frankenstein succeeding in creating a living
being that was larger, stronger and more nimble than any man and rather
intelligent also.  His "monster" had no hatred of people to start
with and was just confused and scared.  However rather than taking
responsibility for his creation, Frankenstein tried to hide from what
he had done and just left his creation roam around free in the Swiss
wilderness to be tomented by various people until his creation
eventually turned into something rather nasty.  When Frankenstein
finally decided that his monster needed to be stopped, well it was
a bit late then as he was really no match for his own creation...
Shane

Kevin Copple wrote:
Here's an idea (if it hasn't already been considered before):

In each executing component of an AGI has Control Code.  This Control Code
monitors a Big Red Switch.  If the switch is turned on, execution proceeds
normally.  If not, all execution by the component stopped.
The big red switch could be a heart monitor on Ben's chest.  The AGI better
keep Ben alive!  And if Ben gets angry with the AGI, he can remove the heart
monitor, so the AGI better keep Ben happy also.
Several other features could be thrown in:
1. Components will check that calling components have the active Control
Code that every component is expected to have.
2. The Control Code checks that certain parameters are not exceeded, such as
AGI memory storage and computational resources.
3. The Big Red Switch monitors not only Ben's heart, but other control
parameters as well.
4. The Big Red Switch monitors another AGI, which has a focused purpose of
looking for trouble with the first AGI.
I imagine that the control code would require a tiny slice of the AGI
resources, yet still be effective.  Implementation details would naturally
require a lot of thought and effort, but I think the concept of a built-in
"virus" would be effective.  Maybe "immune system" would be a more appealing
term.
Kevin Copple



---
To unsubscribe, change your address, or temporarily deactivate your subscription, 
please go to http://v2.listbox.com/member/[EMAIL PROTECTED]



---
To unsubscribe, change your address, or temporarily deactivate your subscription, 
please go to http://v2.listbox.com/member/[EMAIL PROTECTED]

Re: [agi] more interesting stuff

[Shane Legg]

Kevin wrote:
Kevin's random babbling follows:

Is there a working definition of what "complexity" exactly is?  It seems to
be quite subjective to me.  But setting that aside for the moment...
I think the situation is similar to that with the concept of
"intelligence" in the sense that it means different things
to different people.  Indeed there are many formal definitions
of various kinds of complexity, each measuring different things.
For example, is a book on number theory complex?  Well in the everyday
sense of the word most people would say that it is.  In the Kolmogorov
sense it's actually quite simple as the theorems all follow from a
small set of axioms and hence it is quite compressible.  In the logical
depth sense (that is, how many computation cycles would be required to
reproduce the work) the complexity is quite high.  Another example
might be breaking prime number based encryption systems -- there is
little information complexity but a lot of computation time complexity.
Anyway, something can be very complex in one sense while being very
simple in another.  This just seems to show that our intuitive vague
notion of complexity seems to cover a number of very loosely related
things.  Which is a problem when people write about complexity without
being precise about what kind of complexity they are taking about.
Basically they could be talking about just about anything, which is
one of the reasons that I gave up reading non-mathematical papers
on complexity and systems.

Strictly speaking, the noumenal and the phenomenal cannot be separated or
thought of distinctly IMO.  From this viewpoint, complexity is merely
apparent and not fundamentally "real complexity"...
I'd agree to an extent.  Though if all viewpoints have some common
deep down underlying basis of reference such as Turing computation
then perhaps this is THE ultimate viewpoint with which to view the
problem?  This is essentially the approach Kolmogorov compelexity
takes.  As we move away from UTMs towards complex AGI systems the
apparent complexity of things will of course start to change in the
sense that you suggest.  However many of the more fundamental kinds
of complexity (like Kolmogorov and logical depth above) will still
hold firm for very complex things.  In other words: just because
your AGI is really smart doesn't mean that it can now do a large NP
hard problem in a small amount of time --- the problem is likely
to remain very "complex" in some sense (excluding weird stuff like
finding new laws of physics and computation etc... of course).
Cheers
Shane
---
To unsubscribe, change your address, or temporarily deactivate your subscription, 
please go to http://v2.listbox.com/member/[EMAIL PROTECTED]

Re: [agi] A probabilistic/algorithmic puzzle...

[Shane Legg]

Hi Cliff and others,

As I came up with this kind of a test perhaps I should
say a few things about its motivation...

The problem was that the Webmind system had a number of
proposed reasoning systems and it wasn't clear which was
the best.  Essentially the reasoning systems took as input
a whole lot of data like:

Fluffy is a Cat
Snuggles is a Cat
Tweety is a Bird
Cats are animals
Cats are mamals
Cats are dogs

and so on...  This data might have errors, it might be
very bias in its sample of the outside world, it might
contain contradictions and so on... nevertheless we
would expect some basic level of consistency to it.

The reasoning systems take this and come up with all
sorts of conclusions like: Fluffy is an animal based
on the fact that Fluffy is a Cat and Cats seem to be
animals...  In a sense the reasoning system is trying
to "fill in the gaps" in our data by looking at the
data it has and drawing simple conclusions.

So what I wanted to do is to some how artificially
generate test sets that I could use to automatically
test the systems against each other.  I would vary the
number of entities in the space (Fluffy, Cat, Bird...)
the amount of noise in the data set, the number of
data points and so on...

Now the problem is that you can't just randomly generate
any old data points, you actually need at least some kind
of consistency which is a bit tricky when you have some
A's being B's and most B's being C's and all B's not being
D's but all D's being A's.  Before long your data is totally
self contradictorary are are basically just feeding your
reasoning system complete junk and so it isn't a very
interesting test of the system's ability.

So my idea was basically to create a virtual Venn diagram
using randomly placed rectangles as the sets used to compute
the probability for each entity in the space and the conditional
probabilities of their various intersections.  This way your
fundamental underlying system has consistent probabilities
which is a good start.  You can then randomly sample points
from the space or directly compute the probabilities from the
rectange areas (actually n dimensional rectanges as this gives
more interesting intersection possibilities) and so on to get
your data sets.  You can then look at how well the system is
able to approximate the true probabilities based on the
incomplete data that it has been given (you can compute the
true probabilities directly as you know the recatangle areas).

I think I proposed about 6 or so basic variations on this
theme to test the reasoning system's ability with deal with
various level or noise and missing data...  you can come up
with all sorts of interesting variations with a bit of thought.

Yeah, just a fancy Venn diagram really used to generate
reasonably consistent data sets.

Cheers
Shane

---
To unsubscribe, change your address, or temporarily deactivate your subscription, 
please go to http://v2.listbox.com/member/?[EMAIL PROTECTED]

Re: [agi] AIXI and Solomonoff induction

[Shane Legg]

Hi Cliff,

Sorry about the delay... I've been out sailing watching the America's
Cup racing --- just a pity my team keeps losing to the damn Swiss! :(

Anyway:

Cliff Stabbert wrote:


SL> This seems to be problematic to me.  For example, a random string
SL> generated by coin flips is not compressible at all so would you
SL> say that it's alive?

No, although it does take something living to flip the coins; but
presumably it's non-random (physically predictable by observing
externals) from the moment the coin has been flipped.  The decision to
call heads or tails however is not at all as *easily* physically
predictable, perhaps that's what I'm getting at.  But I understand
your point about compressibility (expanded below).


Well I could always build a machine to flip coins...  Or pulls lottery
balls out of a spinning drum for that matter.

Is such a thing predictable, at least in theory?  I have read about
this sort of thing before but to be perfectly honest I don't recall
the details... perhaps the Heisenberg principle makes it impossible
even in theory.  You would need to ask a quantum physicist I suppose.



more and more quickly: the tides are more predictable than the
behaviour of an ant, the ants are more predictable than a wolf, the
wolves are more predictable than a human in 800 B.C., and the human in
800 B.C. is more predictable than the human in 2003 A.D.

In that sense, Singularity Theory seems to be a statement of the
development of life's (Kolmogorov?) complexity over time.


Well it's hard to say actually.  An ant is less complex than a human,
but an ant really only makes sense in the context of the nest that it
belongs to and, if I remember correctly, the total neural mass of some
ants nests is about the same as that of a human brain.  Also whales
have much larger brains than humans and so are perhaps more complex
in some physical sense at least.

A lot of people in complexity believed that there was an evolutionary
pressure driving system to become more complex.  As far as I know there
aren't any particularly good results in this direction -- though I don't
exactly follow it much.

Cheers
Shane

---
To unsubscribe, change your address, or temporarily deactivate your subscription, 
please go to http://v2.listbox.com/member/?[EMAIL PROTECTED]

Re: [agi] AIXI and Solomonoff induction

[Shane Legg]

The other text book that I know is by Cristian S. Calude, the Prof. of
complexity theory that I studied under here in New Zealand.  A new
version of this book just recently came out.  Going by the last version,
the book will be somewhat more terse than the Li and Vitanyi book and
thus more appropriate for professional mathematicans who are used to
that sort of style.  The Li and Vitanyi book is also a lot broader in
its content thus for you I'd recommend the Li and Vitanyi book which
is without doubt THE book in the field, as James already pointed out.

There should be a new verson (third edition) of Li and Vitanyi sometime
this year which will be interesting.  Li and Vitanyi have also written
quite a few introductions to the basics of the field many of which you
should be able to find on the internet.

Cheers
Shane


The Li and Vitanyi book is actually intended to be a graduate-level text in
theoretical computer science (or so it says on the cover) and is formatted
like a math textbook.  It assumes little and pretty much starts from the
beginning of the field; you should have no problems accessing the content.

It is a well-written book, which is a good thing since it is sort of THE
text for the field with few other choices.

Cheers,

-James Rogers
 [EMAIL PROTECTED]

---
To unsubscribe, change your address, or temporarily deactivate your subscription, 
please go to http://v2.listbox.com/member/?[EMAIL PROTECTED]



---
To unsubscribe, change your address, or temporarily deactivate your subscription, 
please go to http://v2.listbox.com/member/?[EMAIL PROTECTED]

Re: [agi] AIXI and Solomonoff induction

[Shane Legg]

Cliff Stabbert wrote:


[On a side note, I'm curious whether and if so, how, lossy compression
might relate.  It would seem that in a number of cases a simpler
algorithm than expresses exactly the behaviour could be valuable in
that it expresses 95% of the behaviour of the environment being
studied -- and if such an algorithm can be derived at far lower cost
in a certain case, it would be worth it.  Are issues like this
addressed in the AIXI model or does it all deal with perfect
prediction?]


Yes, stuff like this comes up a lot in MDL work which can be viewed
as a computable approximation to Solomonoff induction.  Perhaps at
some point a more computable version of AIXItl might exist that is
similar in this sense.

Some results do exist on the relationship between Kolmogorov complexity
and lossy compression but I can't remember much about it off the top of
my head (I'm only just getting back into the whole area myself after a
number of years doing other things!)



What I'm getting at is an attempt at an external definition or at
least telltale of conscious behaviour as either "that which is not
compressible" or "that which although apparently compressible for some
period, suddenly changes later" or perhaps "that which is not
compressible to less than X% of the original data" where X is some
largeish number like 60-90.


This seems to be problematic to me.  For example, a random string
generated by coin flips is not compressible at all so would you
say that it's alive?  Back in the mid 90's when complexity theory
was cool for a while after chaos theory there was a lot of talk
about "the edge of chaos".  One way to look at this is to say that
alive systems seem to have some kind of a fundamental balance between
randomness and extreme compressibility.  To me this seems obvious and
I have a few ideas on the matter.  Many others investigated the subject
but as far as I know never got anywhere.

Chaitin, one of the founders of Kolmogorov complexity theory did
some similar work some time ago,

http://citeseer.nj.nec.com/chaitin79toward.html



The reason I'm thinking in these terms is because I suspected Ockham's
razor to relate to the compressibility idea as you stated; and I've


Sounds to me like you need to read Li and Vitanyi's book on
Kolmogorov complexity theory :)

http://www.cwi.nl/~paulv/kolmogorov.html

Cheers
Shane

---
To unsubscribe, change your address, or temporarily deactivate your subscription, 
please go to http://v2.listbox.com/member/?[EMAIL PROTECTED]

Re: [agi] Breaking AIXI-tl

[Shane Legg]

Eliezer S. Yudkowsky wrote:


Has the problem been thought up just in the sense of "What happens when 
two AIXIs meet?" or in the formalizable sense of "Here's a computational 
challenge C on which a tl-bounded human upload outperforms AIXI-tl?"

I don't know of anybody else considering "human upload" vs. AIXI.

Cheers
Shane

---
To unsubscribe, change your address, or temporarily deactivate your subscription, 
please go to http://v2.listbox.com/member/?[EMAIL PROTECTED]

Re: [agi] Breaking AIXI-tl

[Shane Legg]

Eliezer,

Yes, this is a clever argument.  This problem with AIXI has been
thought up before but only appears, at least as far as I know, in
material that is currently unpublished.  I don't know if anybody
has analysed the problem in detail as yet... but it certainly is
a very interesting question to think about:

What happens when two super intelligent AIXI's meet?

I'll have to think about this for a while before I reply.

Also, you mentioned that it was in your opinion trivial to see that
an AIXI type system would turn into an unfriendly AI.  I'm still
interested to see this argument spelled out, especially if you think
it's a relatively simple argument.

Cheers
Shane


Eliezer S. Yudkowsky wrote:

Okay, let's see, I promised:

An intuitively fair, physically realizable challenge, with important 
real-world analogues, formalizable as a computation which can be fed 
either a tl-bounded uploaded human or an AIXI-tl, for which the human 
enjoys greater success measured strictly by total reward over time, due 
to the superior strategy employed by that human as the result of 
rational reasoning of a type not accessible to AIXI-tl.

Roughly speaking:

A (selfish) human upload can engage in complex cooperative strategies 
with an exact (selfish) clone, and this ability is not accessible to 
AIXI-tl, since AIXI-tl itself is not tl-bounded and therefore cannot be 
simulated by AIXI-tl, nor does AIXI-tl have any means of abstractly 
representing the concept "a copy of myself".  Similarly, AIXI is not 
computable and therefore cannot be simulated by AIXI.  Thus both AIXI 
and AIXI-tl break down in dealing with a physical environment that 
contains one or more copies of them.  You might say that AIXI and 
AIXI-tl can both do anything except recognize themselves in a mirror.

The simplest case is the one-shot Prisoner's Dilemna against your own 
exact clone.  It's pretty easy to formalize this challenge as a 
computation that accepts either a human upload or an AIXI-tl.  This 
obviously breaks the AIXI-tl formalism.  Does it break AIXI-tl?  This 
question is more complex than you might think.  For simple problems, 
there's a nonobvious way for AIXI-tl to stumble onto incorrect 
hypotheses which imply cooperative strategies, such that these 
hypotheses are stable under the further evidence then received.  I would 
expect there to be classes of complex cooperative problems in which the 
chaotic attractor AIXI-tl converges to is suboptimal, but I have not 
proved it.  It is definitely true that the physical problem breaks the 
AIXI formalism and that a human upload can straightforwardly converge to 
optimal cooperative strategies based on a model of reality which is more 
correct than any AIXI-tl is capable of achieving.

Ultimately AIXI's decision process breaks down in our physical universe 
because AIXI models an environmental reality with which it interacts, 
instead of modeling a naturalistic reality within which it is embedded. 
It's one of two major formal differences between AIXI's foundations and 
Novamente's.  Unfortunately there is a third foundational difference 
between AIXI and a Friendly AI.



---
To unsubscribe, change your address, or temporarily deactivate your subscription, 
please go to http://v2.listbox.com/member/?[EMAIL PROTECTED]

Re: [agi] unFriendly AIXI... and Novamente?

[Shane Legg]

Eliezer,

I suppose my position is similar to Ben's in that I'm more worried
about working out the theory of AI than about morality because until
I have a reasonable idea of how an AI is going to actually work I
don't see how I can productively think about something as abstract
as AI morality.

I do however agree that it's likely to be very important, in fact
one of the most important things for humanity to come to terms with
in the not too distant future.  Thus I am open to be convinced that
productive thinking in this area is possible in the absence of any
specific and clearly correct designs for AGI and super intelligence.



crosses the cognitive threshold of superintelligence it takes actions 
which wipe out the human species as a side effect.

AIXI, which is a completely defined formal system, definitely undergoes 
a failure of exactly this type.

Why?

As I see it an AIXI system only really cares about one thing: Getting
lots of reward signals.  It doesn't have an ego, or... well, anything
human really; all it cares about is its reward signals.  Anything else
that it does is secondary and is really only an action aimed towards
some intermediate goal which, in the longer term, will produce yet
more reward signals.  Thus the AI will only care about taking over the
world if it thinks that doing so is the best path towards getting more
reward signals.  In which case: Why take over the world to get more
reward signals when you are a computer program and could just hack
your own reward system code?  Surely the latter would be much easier?
Kind of "AI drugs" I suppose you could say.  Surely for a super smart
AI this wouldn't be all that hard to do either.

I raised this possibility with Marcus a while back and his reply was
that an AI probably wouldn't for the same reason that people generally
say away from hard drugs: they have terrible long term consequences.
Which is to say that we think that the the short term pleasure produced
by the drugs will be out weighed by the longer term pain that results.

However for an AI I think the picture is slightly different as it
wouldn't have a body which would get sick or damanged or die like a
person does.  The consequences for an AI just don't seem to be as bad
as for a human.  The only risk that I can see for the computer is that
somebody might not like having a spaced out computer and then shut it
down and reinstall the system or what ever, i.e. "kill" the AI.  That
wouldn't be optimal for the AI as it would reduce its expected future
reward signal: dead AI's don't get reward signals.  Thus the AI will
want to do what ever it needs to do to survive in the future in order
to maximise its expected future reward signal.

This secondary goal, the goal to survive, could come into conflict
with our goals especially if we are seen as a threat to the AI.

Like I said, reasoning about this sort of thing is tricky so I'm not
overly confident that my arguments are correct...

Cheers
Shane


---
To unsubscribe, change your address, or temporarily deactivate your subscription, 
please go to http://v2.listbox.com/member/?[EMAIL PROTECTED]

Re: [agi] AIXI and Solomonoff induction

[Shane Legg]

Hi Cliff,


So "Solomonoff induction", whatever that precisely is, depends on a
somehow compressible universe.  Do the AIXI theorems *prove* something
along those lines about our universe,


AIXI and related work does not prove that our universe is compressible.
Nor do they need to.  The sun seems to come up most days, the text in
this email is clearly compressible, laws of chemistry, biology, physics,
economics and so on seem to work.  So in short: our universe is VASTLY
compressible.



or do they *assume* a
compressible universe (i.e. do they state "IF the universe is somehow
compressible, these algorithms (given infinite resources) can figure
out how)?


They assume that the environment (or universe) that they have to deal
with is compressible.  If it wasn't they (and indeed any computer based
AI system) would be stuffed.  However that's not a problem as the real
world is clearly compressible...



Assuming the latter, does that mean that there is a "mathematical
definition of 'pattern'"?  As I stated I'm not a math head, but with
what little knowledge I have I find it hard to imagine "pattern" as a
definable entity, somehow.


Yes, there is a mathematical definition of 'pattern' (in fact there
are a few but I'll just talk about the one that is important here).
It comes from Kolmogorov complexity theory and is actually quite
simple.  Essentially it says that something is a pattern if it has
an effective description (i.e. computer program for a Turing machine)
that is significantly shorter than just describing the thing in full
bit by bit.  So for example:

For x = 1 to 1,000,000,000,000
   print "1"
Next

Describes a string of a trillion 1's.  The description (i.e. the
lenght of this program above in bits) is vastly shorter than a
trillion and so a string of a trillion 1's is highly compressible
and has a strong pattern.

On the other hand if I flipped a coin a trillion times and used
that to generate a string of 0's and 1's, it would be exceedingly
unlikely that the resulting string would have any description much
shorter than just listing the whole thing "01000010010010101..."
Thus this is not compressible and has no pattern -- it's random.

There is a bit more to the story than that but not a lot more.



OK, let's say you reward it for winning during the first 100 games,
then punish it for winning / reward it for losing during the next 100,
reward it for winning the next 100, etc.  Can it perceive that pattern?


Clearly this pattern is computationally expressible as so it's no
problem at all.  Of course it will take the AI a while to work out
the rules of the game and on game 101 it will be surprised to be
punished for winning.  And probably for games 102 and a few more.
After a while it will lose a game and realise that it needs to start
losing games.  At game 201 is will probably again get a surprise
when it's punished for losing and will take a few games to realise
that it needs to start winning again.  By game 301 is will suspect
that it need to start losing again and will switch over very quickly.
By game 401 it would probably switch automatically as it will see
the pattern.  Essentially this is just another rule in the game.

Of course these are not exact numbers, I'm just giving you an idea
of what would in fact happen if you had an AIXI system.



Given infinite resources, could it determine that I am deciding to
punish or reward a win based on a pseudo-random (65536-cyclic or
whatever it's called) random number generator?


Yes.  It's "pseudo-random" and thus computationally expressible
and so again it's no problem for AIXI.  In fact AIXItl would
solve this just fine with only finite resources.



And if the "compressibility of the Universe" is an assumption, is
there a way we might want to clarify such an assumption, i.e., aren't
there numerical values that attach to the *likelihood* of gravity
suddenly reversing direction; numerical values attaching to the
likelihood of physical phenomena which spontaneously negate like the
chess-reward pattern; etc.?


This depends on your view of statistics and probability.  I'm a
Bayesian and so I'd say that these things depend on your prior
and how much evidence you have.  Clearly the evidence that gravity
stays the same is rather large and so the probability that it's
going to flip is extremely super hyper low and the prior doesn't
matter to much...



In fact -- would the chess-reward pattern's unpredictability *itself*
be an indication of life?  I.e., doesn't Ockham's razor fail in the
case of, and possibly *only* in the case of, conscious beings*?


I don't see what you are getting at here.  You might need to explain
some more. (I understand Ockham's razor, you don't need to explain
that part; actually it comes up a lot in the theory behind Solomonoff
induction and AIXI...)

Thanks for your comments.

Cheers
Shane


---
To unsubscribe, change your address, or temporarily deactivate your subscription, 
please go to http://v2.listbox.com/

Re: [agi] AIXI and Solomonoff induction

[Shane Legg]

Hi Cliff,


I'm not good at math -- I can't follow the AIXI materials and I don't
know what Solomonoff induction is.  So it's unclear to me how a
certain goal is mathematically defined in this uncertain, fuzzy
universe. 

In AIXI you don't really define a "goal" as such.  Rather you have
an agent (the AI) that interacts with a world and as part of that
interaction the agent gets occasional reward signals.  The agent's
job is to maximise the amount of reward it gets.

So, if the environment contains me and I show the AI chess positions
and interpret its outputs as being moves that the AI wants to make
and then give the AI reward when ever it wins... then you could say
that the "goal" of the system is to win at chess.

Equally we could also mathematically define the relationship between
the input data, output data and the reward signal for the AI.  This
would be a mathematically defined environment and again we could
interpret part of this as being the "goal".

Clearly the relationship between the input data, the output data and
the reward signal has to be in some sense computable for such a system
to work (I say "in some sense" as the environment doesn't have to be
deterministic it just has to have computaionally compressible
regularities).  That might see restrictive but if it wasn't the case
then AI on a computer is simply impossible as there would be no
computationally expressible solution anyway.  It's also pretty clear
that the world that we live in does have a lot of computationally
expressible regularities.



What I'm assuming, at this point, is that AIXI and Solomonoff
induction depend on operation in a "somehow predictable" universe -- a
universe with some degree of entropy, so that its data is to some
extent "compressible".  Is that more or less correct?


Yes, if the universe is not "somehow predicatble" in the sense of
being "compressible" then the AI will be screwed.  It doesn't have
to be prefectly predictable; it just can't be random noise.



And in that case, "goals" can be defined by feedback given to the
system, because the desired behaviour patterns it induces from the
feedback *predictably* lead to the desired outcomes, more or less?


Yeah.



I'd appreciate if someone could tell me if I'm right or wrong on this,
or point me to some plain english resources on these issues, should
they exist.  Thanks.


The work is very new and there aren't, as far as I know, alternate
texts on the subject, just Marcus Hutter's various papers.
I am planning on writing a very simple introduction to Solomonoff
Induction and AIXI before too long that leaves out a lot of the
maths and concentrates on the key concepts.  Aside from being a good
warm up before I start working with Marcus soon, I think it could
be useful as I feel that the real significance of his work is being
missed by a lot of people out there due to all the math involved.

Marcus has mentioned that he might write a book about the subject
at some time but seemed to feel that the area needed more time to
mature before then as there is still a lot of work to be done and
important questions to explore... some of which I am going to be
working on :)



I should add, the example you gave is what raised my questions: it
seems to me an essentially untrainable case because it presents a
*non-repeatable* scenario.


In what sense is it untrainable?  The system learns to win at chess.
It then start getting punished for winning and switches to losing.
I don't see what the problem is.



If I were to give to an AGI a 1,000-page book, and on the first 672
pages was written the word "Not", it may predict that on the 673d page
will be the word "Not.".  But I could choose to make that page blank,
and in that scenario, as in the above, I don't see how any algorithm,
no matter how clever, could make that prediction (unless it included
my realtime brainscans, etc.)


Yep, even an AIXI super AGI isn't going to be psychic.  The thing is
that you can never be 100% certain based on finite evidence.  This is
a central problem with induction.  Perhaps in ten seconds gravity will
suddernly reverse and start to repel rather than attract.  Perhaps
gravity as we know it is just a physical law that only holds for the
first 13.7 billion years of the universe and then reverses?  It seems
very very very unlikely, but we are not 100% certain that it won't
happen.

Cheers
Shane



---
To unsubscribe, change your address, or temporarily deactivate your subscription, 
please go to http://v2.listbox.com/member/?[EMAIL PROTECTED]

Re: [agi] unFriendly AIXI

[Shane Legg]

Hi all,

Some of the discussion seems to take the tone of Novamente vs.
AIXI/AIXI-tl in practical terms.  This simply doesn't make much
sense as AIXI is just a theoretical model.  It would be meaningful
to ask:

a) Is AIXI a good theoretical model of what it is to be an AGI?  i.e.
does it properly define the elusive concept of "general intelligence"
and does the AIXI system demonstrate that at some theoretical level
such a thing is both logically consistent and amazingly powerful?
Is it good enough to act as the very definition of what it is to be
a super powerful AGI?  If the answer is "no" then why?  If the answer
is "yes" then in AIXI we have a precise mathematical definition of
what super AGI is in which case;

b) How good a practical system is Novamente when viewed from the
perspective of the AIXI theoretical model of AGI?


There have been some comments made about AIXI having a fixed goal.
As Eliezer pointed out, there isn't some fixed goal built into
AIXI, it's just a part of the system's analysis:  *if* you had
some kind of fixed goal in mind and just threw the AIXI system
at it how would it cope?  ... and the answer is that it would
quickly work out what the goal was and then kick butt.

However even within this scenario the concept of "fixed goal" is
something that we need to be careful about.  The only real goal
of the AIXI system is to get as much reward as possible from its
environment.  A "goal" is just our description of what that means.
If the AI gets reward for winning at chess then quickly it will get
very good at chess.  If it then starts getting punished for winning
it will then quickly switch to losing at chess.  Has the goal of
the system changed?  Perhaps not.  Perhaps the goal always was:
Win at chess up to point x in time and then switch to losing.
So we could say that the goal was always fixed, it's just that up
to point x in time the AI thought the goal was to alway win and it
wasn't until after point x in time that it realised that the real
goal was actually slightly more complex.  In which case does it make
any sense to talk about AIXI as being limited by having fixed goals?
I think not.


Ben often says that AIXI isn't really that big a deal because it's
trivial to build a super powerful AI given infinite computational
resources.  However to the best of my knowledge, Hutter's was the
first to actually do this and properly analyse the results: AIXI
is the first precisely defined and provably super powerful fully
general theoretical AI to be proposed and carefully analysed.
(Solomonoff only dealt with the more limited sequence prediction
problem)  In which case it seems, at least to me, that Hutter has
done something quite significant in terms of the theory of AI.
Is Hutter's work of practical use?  Well that's an open question
and only time will tell.


Finally about AIXI/AIXItl needing infinite resources.  AIXI contains
an uncomputable function so that's the end of that.  AIXItl however
only requires a finite amount of resources to solve any given problem.
However in general AIXItl requires unbounded resource when we consider
its ability to face problems of unbounded complexity.  Clearly this
will be true of any system that is able to effectively deal with
arbitrary problems -- it will require arbitrary resources.  This is
not some special property of AIXItl, all systems in this class must
be like this, in particular if Novemente is able to solve arbitrary
problems then it must too have access to arbitrary resources.

The flip side is that if we restrict the complexity of the problems
that AIXItl faces to be finite (in an appropriate way), then it only
ever requires finite resources just like Novamente.  The point of all
this is this:  The finiteness or lack of finiteness that AIXItl
requires is nothing especially bad in the sense that no system can do
better in these terms.  Thus talking about AIXItl being irrelevant
because it requires infinite resources doesn't make any sense as
Novamente could not do any better in terms of requiring unbounded
resources when faced with arbitrary problems.  You can however argue
that the way in which AIXItl requires resources as the complexity of
a problem increases is highly non-optimal compared to Novamente.
Perhaps this is true, however the purpose of AIXItl is as a theoretical
tool to study this class of AI's and their properties.  If we could
prove that AIXItl or a variant or even a completely different model
was optimal in this latter sense also then that would be very cool
indeed and another step towards a truly practical super AI.

Ok... email is getting too long now

bye

:)

Shane

---
To unsubscribe, change your address, or temporarily deactivate your subscription, 
please go to http://v2.listbox.com/member/?[EMAIL PROTECTED]

Re: [agi] Godel and AIXI

[Shane Legg]

Which is more or less why I figured you weren't going to "do
a Penrose" on us as you would then fact the usual reply...

Which begs the million dollar question:

Just what is this cunning problem that you have in mind?

:)

Shane

Eliezer S. Yudkowsky wrote:

Shane Legg wrote:


Eliezer S. Yudkowsky wrote:


An intuitively fair, physically realizable challenge with important 
real-world analogues, solvable by the use of rational cognitive 
reasoning inaccessible to AIXI-tl, with success strictly defined by 
reward (not a Friendliness-related issue).  It wouldn't be 
interesting otherwise.


Give the AIXI a series mathematical hypotheses some of which are
Godelian like statements and asking the AIXI it if each statement
is true and then rewarding it for each correct answer?

I'm just guessing here... this seems too Penrose like, I suppose
you have something quite different?



Indeed.

Godel's Theorem is widely misunderstood.  It doesn't show that humans 
can understand mathematical theorems which AIs cannot.  It does not even 
show that there are mathematical truths not provable in the Principia 
Mathematica.

Godel's Theorem actually shows that *if* mathematics and the Principia 
Mathematica are consistent, *then* Godel's statement is true, but not 
provable in the Principia Mathematica.  We don't actually *know* that 
the Principia Mathematica, or mathematics itself, is consistent.  We 
just know we haven't yet run across a contradiction.  The rest is 
induction, not deduction.

The only thing we know is that *if* the Principia is consistent *then* 
Godel's statement is true but not provable in the Principia.  But in 
fact this statement itself can be proved in the Principia.  So there are 
no mathematical truths accessible to human deduction but not machine 
deduction.  Godel's statement is accessible neither to human deduction 
nor machine deduction.

Of course, Godel's statement is accessible to human *induction*.  But it 
is just as accessible to AIXI-tl's induction as well.  Moreover, any 
human reasoning process used to assign perceived truth to mathematical 
theorems, if it is accessible to the combined inductive and deductive 
reasoning of a tl-bounded human, is accessible to the pure inductive 
reasoning of AIXI-tl as well.

In prosaic terms, AIXI-tl would probably induce a Principia-like system 
for the first few theorems you showed it, but as soon as you punished it 
for getting Godel's Statement wrong, AIXI-tl would induce a more complex 
cognitive system, perhaps one based on induction as well as deduction, 
that assigned truth to Godel's statement.  At the limit AIXI-tl would 
induce whatever algorithm represented the physically realized 
computation you were using to invent and assign truth to Godel 
statements.  Or to be more precise, AIXI-tl would induce the algorithm 
the problem designer used to assign "truth" to mathematical theorems; 
perfectly if the problem designer is tl-bounded or imitable by a 
tl-bounded process; otherwise at least as well as any tl-bounded human 
could from a similar pattern of rewards.

Actually, humans probably aren't really all that good at spot-reading 
Godel statements.  If you get tossed a series of Godel statements and 
you learned to decode the diagonalization involved, so that you could 
see *something* was being diagonalized, then the inductive inertia of 
your success at declaring all those statements true would probably lead 
you to blindly declare the truth of your own unidentified Godel 
statement, thus falsifying it.  Thus I'd expect AIXI-tl to far 
outperform tl-bounded humans on any fair Godel-statement-spotting 
tournament (arranged by AIXI, of course).



---
To unsubscribe, change your address, or temporarily deactivate your subscription, 
please go to http://v2.listbox.com/member/?[EMAIL PROTECTED]

Re: [agi] unFriendly AIXI

[Shane Legg]

Eliezer S. Yudkowsky wrote:


An intuitively fair, physically realizable challenge with important 
real-world analogues, solvable by the use of rational cognitive 
reasoning inaccessible to AIXI-tl, with success strictly defined by 
reward (not a Friendliness-related issue).  It wouldn't be interesting 
otherwise.


Give the AIXI a series mathematical hypotheses some of which are
Godelian like statements and asking the AIXI it if each statement
is true and then rewarding it for each correct answer?

I'm just guessing here... this seems too Penrose like, I suppose
you have something quite different?

Shane

---
To unsubscribe, change your address, or temporarily deactivate your subscription, 
please go to http://v2.listbox.com/member/?[EMAIL PROTECTED]

Re: [agi] A thought.

[Shane Legg]

Mike Deering wrote:


types of ENVIRONMENT are very broad, perhaps infinite.  Just as there is 
no "general" environment, there is no "general" intelligence.  A mind 
must be matched to its environment.

So in essence you are claiming that truly general intelligence
is impossible and thus AGI must also be impossible.

Have I understood you correctly?

Cheers
Shane

---
To unsubscribe, change your address, or temporarily deactivate your subscription, 
please go to http://v2.listbox.com/member/?[EMAIL PROTECTED]

Re: [agi] An Artificial General Intelligence in the Making

[Shane Legg]

Daniel,

An ARFF file is just a collection of n-tuple data items where each tuple
dimension has defined type information.  It also has a dimension that
is marked as being the "class" of the data item.  So because it's
basically just a big table of data you could in theory put any kind of
information you like in there provided that you are a little creative
in the encoding.

However while you could do something like that with an ARFF file it
probably doesn't make much sense.  ARFF files carry with them the
implicit assumption that the data items are more or less i.i.d. and
that you suspect that there is some sort of explicit relationship
between the dimensions; in particular you usually are interested in
the abilty to predict the class dimension using the other dimensions.
This is how Weka classifiers interpret the files.

So in short:  I'm sure you could jam KNOW data into an Arff file but
I don't really see why doing so would make much sense.

Cheers
Shane

Daniel Colonnese wrote:

For those of us who are following the KNOW thread, could somebody 
comment on the capabilities of KNOW beyond existing knowledge 
representation language such as the ARFF format for the popular WEKA system.
 
I've input data into such a system before and while existing systems 
have extensive grammar for representing logical relations they have very 
limited capabilities for more ambiguous  knowledge.
 
The KNOW document Ben posted a link too says:
 
"Syntax to semantics mapping in the natural language module, in which 
the final result should be represented in this language;"
 
This kind of capabilities would certainly be a huge advance over 
something like ARFF.  If anyone works with ARFF, could he or she comment 
on the possibilities of such translation with the ARFF grammar?  Does 
anyone who's familiar with the technical workings of knowledge 
representation language have any idea on how this kind of mapping could 
be accomplished? 
 
-Daniel



---
To unsubscribe, change your address, or temporarily deactivate your subscription, 
please go to http://v2.listbox.com/member/?[EMAIL PROTECTED]

Re: [agi] C-T Thesis (or a version thereof) - Is it useable as anin-principle argument for strong AI?

[Shane Legg]

Hi,

This isn't something that I really know much about, but I'll
put my understanding of the issue down in the hope that if
I'm missing something then somebody will point it out and
I'll learn something :)

The literal Church-Turing thesis states that all formal models
of what constitutes a well defined process are in fact equivalent
to the Turing machine model.  This thesis came about after it
was discovered that all the various formal models (lambda
calculus, recursive function theory and many others) that
had been proposed were provably equivalent to the TM model.
It is worth noting that nobody has actually proven that this
claim is true, it's more the case that all efforts to find
formal model of "well defined processes" that's more powerful
than a Turing machine model have all failed and so people
assume that the thesis probably true.

Some people take this a step further and claim that not only
are all formal models of well defined processes equivalent,
but in fact all well defined physical processes are also
equivalent to the Turing machine model.  This appears to be
supported by the fact that no "well defined" physical process
has ever been found that is more powerful than the Turing
machine model.  Thus in a sense this claim is very similar
to the one above as it essentially rests on empirical evidence
rather than hard proof.

If this physical interpretation of the Church-Turing thesis
is accepted then it follows that if the physical brain and its
operation is a "well defined process" then it must be possible
to implement the process that the brain carries out on a Turing
machine.  This is the claim of "Strong AI."

Does that sounds correct to people?

Cheers
Shane


Anand AI wrote:

Hi everyone,

After having read quite a bit about the the C-T Thesis, and its different
versions, I'm still somewhat confused on whether it's useable as an
in-principle argument for strong AI.  Why or why isn't it useable?  Since I
suspect this is a common question, any good references that you have are
appreciated.  (Incidentally, I've read Copeland's entry on the C-T Thesis in
SEoC (plato.standford.edu).)

I'll edit any answers for SL4's Wiki (http://sl4.org/bin/wiki.pl?HomePage),
and thanks very much in advance.

Best wishes,

Anand


---
To unsubscribe, change your address, or temporarily deactivate your subscription, 
please go to http://v2.listbox.com/member/?[EMAIL PROTECTED]

Re: [agi] AI and computation (was: The Next Wave)

[Shane Legg]

Pei Wang wrote:


Again, I'm not saying that NARS is not a TM in any sense, but that it is not
a TM at the questing-answering level.  As I said in the paper, if you
consider the life-long history of input and output of the whole system, NARS
is a TM.  Also, if you check each individual inference step, NARS is also a
TM. However, when we talk about the "computability" and "computational
complexity" of a problem/question, it is at neither of the two levels (it is
shorter than the whole life, but longer than a single step).  It is at this
level NARS is not a TM (and not merely "doesn't LOOK like a TM").


Ok, I understand this --- that NARS is a TM at both the life-history
level and at the inference step level.  Which is pretty much what I
had in mind.

As for the in between level, I'm going to have to read your paper
some more and think about it and get back to you.

Cheers
Shane

---
To unsubscribe, change your address, or temporarily deactivate your subscription, 
please go to http://v2.listbox.com/member/?[EMAIL PROTECTED]

Re: [agi] AI is a trivial problem

[Shane Legg]

Ben,

We seem to be thinking along similar lines in most aspects here...



The way the human brain seems to work is:

* some of its architecture is oriented towards increasing the sum over n of
int(S,n,rS,rT), where rS is given by brain capacity (enhanced by tools like
writing?) and rT is different for different problems, but is a "humanly
meaningful time scale."

* some of its architecture is oriented towards increasing problem-solving
ability for n that are so large that int(S,n,rS,rT) is totally  miserable
for realistic (rS, rT)


Right, for example my brain doesn't solve visual processing problems
using completely general learning structures.



That is: any real-world-useful general intelligence is going to have a mix
of "truly general intelligence methods" focused on boosting int(S,n,rS,rT)
for small n, and "specialized intelligence methods" that are different from
narrow-AI methods in that they specifically leverage a combination of

* specialized heuristics
* the small-n general intelligence methods


Yes, though we will have to be careful because an AI is likely to be
a lot more dynamic than our brains.  I'm not even talking about fancy
stuff like having the AI recode itself, just the fact that an AI would
be able to solve some problems using very general intelligence and then
adapt that solution into a very efficient and quite specialized ability
with relative easy.  That ability could then be a stepping stone towards
other problems and the development of further more complex specialized
abilities as required.



Now, we're asking not just whether S can solve simple problems, we're asking
whether S can solve problems that are simple against the background of its
environment and its own mind at a given point in time.

In this case, I think the same conclusions as above will hold, but more
weakly.  I.e. the general intelligence capability will hold robustly for
somewhat larger n.  But still there will be the dichotomy between small-n
general intelligence, and large-n specialized-building-on-general
intelligence.  Because only some problems relating to self and environment
are useful for achieving system goals, and the system S will be specialized
for solving *these* problems.


Which I think is more or less what I also said above.  Here the training
of the AI becomes very important as this is what develops the abilities
of the system.  Or put another way; it's often the case that knowing how
to solve a simpler or related problem to the problem you are currently
facing is very useful.  We don't just solve all problems from scratch,
we draw on our experiences with similar problems from the past.



Now, moving on, I'll make the following claim:

** achieving "small-n general intelligence" is a basically simple math
problem **


As you always like to say: given infinite resources... ;)



I think that the hard problem of AGI is actually the other part:

BUILDING A SYSTEM CAPABLE OF SUPPORTING SPECIALIZED INTELLIGENCES THAT
COMBINE NARROW-AI HEURISTICS WITH SMALL-N GENERAL INTELLIGENCE


Yes this is part of the problem, the other thing you don't mention
is the difficulty of trying to solve small-n problems efficiently.



I agree, Shane, that algorithmic information theory is useful for the
"small-n general intelligence" part.  But it's just providing a complicated,
sometimes elegant mathematical formalism for what is actually one of the
*easier* parts of the practical AGI task.


I would say that there is something very important here that you haven't
mentioned:  The value in having a precisely mathematically defined and
provably strong definition of what general intelligence actually is.

Clearly I accept that this is possible and I think you do too.  However
let's not forget that almost NOBODY else in the field of AI (or psych
or anywhere else) accepts this or has even heard of the idea before!
Even if they had heard of it I'm sure there would be an enormous
amount of resistance to the idea.

So yeah, in terms of the "practical AGI task" is might not turn out
to be all that huge a deal, I'm really not sure.  But in terms of
actually getting a sizable number of people all agreeing on what the
hell the goal at least in theory is, it could be very significant.

In the words of Charles Kettering,

"A problem well stated is a problem half solved."

The work of Marcus Hutter is, I believe, currently the most
significant piece of work in this direction to date.

Cheers
Shane


---
To unsubscribe, change your address, or temporarily deactivate your subscription, 
please go to http://v2.listbox.com/member/?[EMAIL PROTECTED]