Re: determinism, freewill, chaos, and circular causality
 From: curt@xxxxxxxx (Curt Welch)
 Date: 26 Apr 2006 02:46:57 GMT
"Stephen Harris" <cyberguard1048@xxxxxxxxx> wrote:
"Curt Welch" <curt@xxxxxxxx> wrote in message
news:20060419151144.187$xS@xxxxxxxxxxxxxxxxx
"Stephen Harris" <cyberguard1048@xxxxxxxxx> wrote:
"Curt Welch" <curt@xxxxxxxx> wrote in message
news:20060417185142.082$bI@xxxxxxxxxxxxxxxxx
"JGCASEY" <jgkjcasey@xxxxxxxxxxxx> wrote:
Curt Welch wrote:A turning machine emulator, running on a turning machine, can't be
...
made more powerful, buy changing the code of the emulator. Anything
you could do by changing the code of the emulator, could also have
been done, by running a program on the emulator. What makes you
think that a learning machine can be made more powerful by allowing
it to change it's own code? If you code it right to start with,
there's nothing new it will be able to learn by changing it's own
code, that it could not have learned by changing
the structures which it was originally set up to evolve through
learning.
It seems to be to be the same as the turning machine problem. What
you would need to do is build a signal processing learning machine
that has enough flexibility to solve any signal processing problem
in a given domain
(time/frequency/complexity). This is an interesting question, and I
don't know if anyone has found a way to prove that any some class of
machines would have the power to solve any possible learning problem
given enough time and memory. It's seems possible to me.

I think you probably meant a Turing machine which solves computable
problems. Marcus Hutter has explored a noncomputable AI approach:
http://www.idsia.ch/~marcus/ Universal Artificial Intelligence (2nd
choice)
"There are strong arguments that AIXI is the most intelligent unbiased
agent possible in the sense that AIXI behaves optimally in any
computable environment. The book outlines for a number of problem
classes, including sequence prediction, strategic games, function
minimization, reinforcement and supervised learning, how they fit into
the general AIXI model and how AIXI formally solves them. The major
drawback of the AIXI model is that it is incomputable. The book also
presents a preliminary computable AI theory. We construct an algorithm
AIXItl, which is superior to any other time t and space l bounded
agent."
SH: The Turing machine is an ideal machine, so has no physical
limitations.
Actual computers are constrained by physical resources and by time in
that there are intractable problems which cannot be computed in the
time available before the universe experiences heat death. Physical
computers have proven computational limitations:
http://www.cis.udel.edu/~case/colt.html
John Case's COLT Page
"Consider the problem of finding a rule for generating a sequence of
numbers such as 9, 61, 52, 63, 94, 46, 18, 1, 121, 441, ... . Here
is a rule for this sequence. First compute the squares of successive
integers beginning with 3, but, then, to generate the sequence, use,
in place of these squares, the squares each with its decimal digits
written down in reverse order (ignoring any lead zeros). N.B. This
rule can be written as a formal algorithm (or computer program).
The problem of finding such rules gets harder as the sequences to
generate get more complicated than the one above. Can the rule
finding itself be done by some computer program? Interestingly,
it is mathematically proven that there can be no computer program
which can eventually find (synonym: learn) these (algorithmic) rules
for all sequences which have such rules!"
That sounds interesting. Do you happen to know what this proof is?
No. I sent him, John Case, an email once asking this question, but did
not receive a satifactory answer. At first I thought it was related to
the Halting problem but now I think the reason is Intractable restraints
on time and resources. Wolfram thinks the universe has evolved from a
cellular automata rule which now has an exceedingly complex structure.
He admits that to test this theory one would have to select a likely
candidate and run that rule for about the same age as the universe in
order to confirm a match, though there would be good indications
after only 9 billion years.
It seems to me that for a turning machine with infinite time and space
to work with, it should be able to find the answer to any finite sized
rule (any finite sized program for generating a sequence.
Well let us take Pi, which is expanded by computing the relationship
of the perimeter to the diameter of a circle. Pi is defined as computable
and passes all randomness tests. A Turing machine can expand Pi
forever, adding one digit at a time. It can produce a sequence of
finite Pi values which increments and never terminates.
There are several algorithms (rules) for generating Pi which are short
and finite. Suppose I extract some finite portion of the infinite
expansion of Pi. Now suppose I have a geiger counter which measures
random decay of the nucleus (or any fair random generation method) and a
scale, perhaps in miliseconds, which records clicks in a range of
09, sort of like rolling a fair tensided di with numbers on each side.
I don't tell you the source of these two sequences of random numbers,
both of which are finitely long. Can you distinguish between the two of
them? I don't think so. Since both are random, somewhere within the
expansion of Pi, the sequence of digits generated by random atomic
decay will appear.
Can you prove that? I mean, just because Pi seems random in all ways we
know to test it, how can you prove the algorithm will actually produce some
finite string of length say 10^100 at some point?
Do the algorithms for generating Pi all require infinite memory? i.e.,
increasingly larger amounts of memory as the number of digits produced
increases?
The finite sequence generated by the rule for Pi
will somewhere include the random atomic sequence. If it doesn't,
then it isn't random.
There are tests for randomness but no proofs because how could
one prove an infinitely long number is infinite. How does one
determine if a short finite sequence originated from a random
process or a process which is not random but produces random
output? Chaitin and others have defined what they call "truly random
numbers" and the qualifcation is that there is no shorter input string
to generate the output, there is no rule or algorithmic process.
That seems intuitive. But does anyone have a way to prove that a given
finite string is truly random?
I don't think there is a way to distinguish between some randomly
extracted subsequence of a truly random number and a randomly
extracted subsequence of Pi if you don't first know the source.
A random number expresses all possible random subsequences
infinitely often.
It seems to me it would be possible to write an algorithm that generated an
infinite string of digits in such a way as to guarantee that at some point.
I guess it's trivial to do since you can just generate all strings of
length 1, then 2, then 3 and just keep going. But that would require
memory equal to the size of the string you were generating. It seems to me
however it might be possible to find such an algorithm that had the
property that you could prove every finite string would be generated at
some point but yet use far less memory then the string you were generating.
Or, looking at it from the other direction, it would imply that for any
finite string you could generate there would aways be an algorithm for
generating it that would be shorter than the string.
This would just imply that the language you used for specifying the
algorithm was simply more dense than the language of digits. Which seems
potentially possible since whatever language you specified the rules in
would in fact have more meaning to each digit than a simple list of digits.
It seems to me that you could not answer these sort of questions without
first putting some limits on the type of language you could use to specify
your rules with. For example, if I could specify a rule like this in my
"rule language":
Generate 10256 digits of Pi starting at digit offset 18237237862373
And the above rule was encoded like this:
111110 10256 111111111111110 18237237862373
Then you have a simple rule for specifying large finite strings in a
language that might be shorter than the strings it generated for most
cases. Or maybe not because the offset value might tend to be larger than
the string you were trying to generate for most cases.
It just seems to me however that the power to specify an algorithm would
tend to allow you to encode any string in a shorter form with a powerful
enough rule language.
I wrote all this to bring up one possibility in reply
to:
It seems to me that for a turning machine with infinite time and space
to work with, it should be able to find the answer to any finite sized
rule (any finite sized program for generating a sequence.
Suppose somebody discovers another number X, like Pi, computable,
infinite and passing all randomness tests. How, by just intercepting
a partial sequence of an encrypted transmission, could you tell if
this sequence originated from Pi, or number X, or a truly random
number, or some other number in the category as X and Pi?
If there is more than one possible answer, it should be able to find them
all.
Ah, but no, it couldn't. Because the list of possible rules is infinite if
you put no limit on the size of algorithm that is generating the string
(there's nothing stopping you from written really bad code). So the best
you could do is specify a program that would find them all, from shortest
to longest, but it would never terminate and never stop finding more
answers.
I am not so sure that there is a finite input rule for the Turing machine
which allows it to discern which of an infinite number of finite or
infinite patterns, some truly random, some like Pi, corresponds to a
given finite sample under examination?
Well sure. You could write a program to generate all possible programs,
and emulate each one to see if it produced the finite sample under
examination. Ah, but there is the halting problem in there that could
prevent you from knowing if the program under examination was every going
to finish. Yeah, there might be issues. You might be able to get around
those issues by specifying a restricted programming language which was
guaranteed to keep producing output digits so you could know that the test
would always terminate.
As a matter of fact, the solution is obvious. Generate all possible
programs and see if they produce the correct sequence. Finite sized
Aren't there an infinite number of these finite programs? Like there
are an infinite number of finite representations of the expansion of Pi.
Pi to a hundred digits, to a million digits, to a zillion digits ...
I was thinking simply that there is a finite number of programs of length
10K bytes or less. It would require to you specify the language you were
using to specify the program.
programs can produce very very large sequences  but they are forced to
repeat. So there's only a finite amount of testing required for each
program to know if it matches. Ah, but no.... You don't know how long
the program is which is generating the sequence so you never know how
much you have to test. No matter how much you test, you might not have
tested enough  so it's impossible to ever complete the test.
Pi is a short rule which never repeats (random) when expanded
indefinitely on a Turing machine. If you mean on a physical computer, I
don't know if it would repeat, but it would run out of resources due to
combinatorial explosion, intractability, when the rule was sufficiently
complex evolved. I don't think there is a computable way to backtrace the
universe to a single cellular automata rule given finite resources. Or
even Minsky says you can't reproduce intelligence by running zillions of
speeded up scenarios on the evolution of life on our planet.
So, I think the only thing you could write a program to test is a
sequence of finite length. Given any sequence of finite length you
could always find a set of rules to produce it (which I guess is
obvious since the program only needs to be "print sequence").
Well, I thought we were discussing printing out (predicting) the next
number in the sequence, not repeating the same input sequence.
We might have started there. I can't remember how we got here. :)
But the problem of prediction is a much cleaner and and easier one to
address than the problem of producing an algorithm to generate a fixed
sized string.
That
is BTW, how they define a truly random number. The input length
and the output length must be of the same length. There is no shorter
rule which exists for generating a longer output length (such as Pi).
Also don't you need to discover the unique rule which generated the
finite output under examination? I'm pretty sure there are an infinite
number of programs that can generate a given finite output. But the
next digit(s) in the given finite sequence (predicting it) can change
with every one of those infinte programs which generated the first
specified finite output. Many rules could generate the sequence
used as an example by John Case, and the next digit could be
different than 169 (using his example rule) with each different rule.
"Consider the problem of finding a rule for generating a sequence of
numbers such as 9, 61, 52, 63, 94, 46, 18, 1, 121, 441, ... . Here
is a rule for this sequence. First compute the squares of successive
integers beginning with 3, but, then, to generate the sequence, use,
in place of these squares, the squares each with its decimal digits
written down in reverse order (ignoring any lead zeros).
And for an infinite sequence
produce by a finite size set of rules, you could never know when to
stop testing the sequence to know if you current set of rules is
actually producing the right sequence.
So, as a practical program you could write a program to test all
possible rule sets to see which ones currently worked to produce the
first N numbers
of the sequence for any N, but not for all N.
The rules may be finitely long, but that doesn't mean there is a finite
number of such rules. I think there are a infinite number of finite rules
for producing any finite sequence.
Yeah, I think that must be true.
I'm not catching your drift any more.
How many rules/programs are there to produce any given number such as 2?
Yeah, I'm drifing for sure. :)
Regards,
Stephen

Curt Welch http://CurtWelch.Com/
curt@xxxxxxxx http://NewsReader.Com/
.
 References:
 determinism, freewill, chaos, and circular causality
 From: feedbackdroid
 Re: determinism, freewill, chaos, and circular causality
 From: Curt Welch
 Re: determinism, freewill, chaos, and circular causality
 From: JGCASEY
 Re: determinism, freewill, chaos, and circular causality
 From: Curt Welch
 Re: determinism, freewill, chaos, and circular causality
 From: Stephen Harris
 Re: determinism, freewill, chaos, and circular causality
 From: Curt Welch
 Re: determinism, freewill, chaos, and circular causality
 From: Stephen Harris
 determinism, freewill, chaos, and circular causality
 Prev by Date: Re: when can bayesian "not" be applied?
 Next by Date: Re: Is a general purpose mechanism possible?
 Previous by thread: Re: determinism, freewill, chaos, and circular causality
 Next by thread: Re: determinism, freewill, chaos, and circular causality
 Index(es):
Relevant Pages
