Re: [agi] A difficulty with AI reflectivity

From: Paul Fidika (fidika@gmail.com)
Date: Thu Oct 21 2004 - 14:38:54 MDT


Well I'm afraid your problem is a bit beyond me at the moment, but
after thinking about it for two days here are my thoughts.

First of all, it may be helpful to rephrase and clarify Eliezer's
problem in more concrete terms. (Disclaimer: the following discussion
makes use of Algorithmic Information theory, a topic which I barely
understand, and as such may be incorrect.) Suppose we have a Seed-AI
and a particular problem which we want the Seed AI to solve for us,
namely:

(P) Find and output (and only output) the first random bit-string of
length n, and then halt,

where by "first" we mean the usual ordering over binary strings, and
by "random" we mean random in the sense of incompressibility; a
bit-string B of length n is random if the Kolmogorov-complexity of the
smallest program which prints B is greater than or equal to n.
(Intuitively this means that the smallest program which prints B is
just the program "print 0011101…", where 0011101… is B.)

Now our problem P is very simple and may be encoded in a fixed number
of bits, so that the complexity of P, that is, K(P), is a fixed
constant. Our Seed AI can also be encoded in a fixed number of bits,
K(S), and n (the n made reference to in P) may be encoded in log(n)
bits. Now choose an n such that

n > K(S) + K(P) + log(n).

Such an n exists since K(S) and K(P) are both constants. Now if our
Seed AI were to output a bit-string as an answer to P, then our Seed
AI would be provably wrong, because, by definition, the bit-string
being asked for by P must be of complexity n or greater, whereas we
have program of complexity less than n which has printed it out,
meaning the string our Seed AI has printed out in fact has complexity
less than n, and hence cannot be the string requested in P. Thus we
have constructed a problem which is IMPOSSIBLE for our Seed AI to
solve, NO MATTER WHAT. Our Seed AI may shuffle around its bits all it
wants, it will be to no avail.

Notice that this construction works for ANY computable-intelligence;
given any computable-intelligence, we can ALWAYS construct a problem
which this intelligence provably cannot solve. (Yes Penrose, that
INCLUDES humans.) A few properties of P are worth noting:

-There are a finite number of candidate-solutions to P (2^n, to be exact).
-A unique solution exists.
-If we weaken P to "print any (but only) one random binary string of
length n", then most of the candidate-solutions are in fact solutions
to this weakened P (most strings of length n are random)
-Whatever bit-string the Seed AI returns, it will provably be the incorrect one.
-P can be thought of as a type of Gödel sentence.

The problem is that our Seed AI is of fixed complexity, and cannot
increase its complexity by internally rewriting its program (though it
can decrease its complexity), and its current complexity is not
sufficient to solve P. What if however our Seed AI turns to the
environment to gain complexity? Suppose that our environment contains
some complicated mechanism which proposes possible changes to the Seed
AI architecture; the Seed AI reviews the proposed architecture, and
then either decides (in a finite amount of time) whether or not to
change to the suggested architecture or stay the way it is. One such
proposed architecture-change must be powerful enough to solve a given
instance of P, so all our Seed AI needs to do is be capable of
recognizing such an architecture when it sees one (it needn't be able
to generate that architecture internally). Suppose that whenever our
Seed AI accepts an architecture change when considering solving an
instance of P, then this new architecture can indeed solve P (note
that the converse needn't hold—our Seed AI may reject architectures
which would have otherwise solved the instance of P, and the following
argument holds even if you insert a finite number of meta-architecture
changes, that is you propose changes to the Seed AI which will enable
it to better recognize architectures useful for better recognizing
architectures useful for solving P, and so on).

If our Seed AI did accept some such architecture change, changed to
the new architecture, and correctly printed out the string requested
by P, then, prima facie, there appears to be no contradiction, because
K(S) + K(A), where A is the new architecture, may be much larger than
n. However, then we can create a program G which takes our Seed AI and
proposes to it EVERY possible architecture-change in some standard
order of enumeration, and then, if the Seed AI accepts a new
architecture, this new architecture is run and presumably prints the
string requested by P. But this program can also be specified in K(G)
bits, so for all n > K(S) + K(G) + K(P) + log(n) not only can our Seed
AI not solve P, our Seed AI is incapable of recognizing an
architecture which CAN solve P when it sees one.

Also, Marc Geddes's suggestion of weakening our requirements from "the
Seed AI outputs the string specified by P" to "the Seed AI outputs a
string which is probably the string specified by P" does not help at
all. If our Seed AI approximates the Kolmogorov-Complexity function
for all strings of length n (this is possible—though the
Kolmogorov-complexity function is uncomputable, it is semi-computable
from above), and, after a certain amount of time (as determined by the
AI), the Seed AI halts and outputs the string which is most likely to
be the string requested by P as determined by the
approximation-thus-far, then one would think that the Seed AI's answer
would become "more likely correct" if it spent a longer amount of time
approximating the Kolmogorov-Complexity function. But in fact whatever
string the Seed AI outputs will always be PROVABLY wrong, for the same
reasons mentioned above.

(Note: If however the time at which the Seed AI halts and outputs is
determined by some external mechanism of complexity greater than the
difference between n and K(S) + K(P) + log(n), it might be possible
for the Seed AI to output the correct string. Furthermore, if the Seed
AI is given a large enough random string as input initially, then it
is possible for the Seed AI to solve P by transforming this string
into a string of length n and outputting it, which might be the
correct answer. I'm not sure if, no matter how much random-input we
give the Seed AI, the Seed AI can do much better than this, that is,
much better than random-guessing, for an arbitrary instance of P. Any
ideas?)

Eliezer Wrote:

> The unique, new problem comes when we ask the theorem prover to rewrite
> itself entirely. Even if we adjoin to Theory-1 the assumption that
> Theory-0 is consistent, if a Theory-1 prover were to write a provably
> consistent (in Theory-1) prover, it would write a Theory-0 prover. This
> prover would then be unable to approve any further rewrites.
>
> We may be able to rescue Schmidhuber's Godel Machine by compartmentalizing
> it into an object system that provably has expected utility for solving
> problems, and a meta-system that can only be rewritten if the rewrite
> provably admits only theorems admissable in the original meta-system. That
> is, we use *two different* criteria for modifying *two different*
> components of the Godel Machine. I don't regard this as a good solution to
> the deep AGI problem [...] It is furthermore unclear as to how a
> rewrite of the meta-system would be adjudged *superior* to the prior
> system, even if it were proven admissible.

Your problem appears to be that the Seed AI can never (internally)
increase its complexity, whereas it can decrease it. Also, the Seed AI
will not, in general, have sufficient complexity to recognize an
improvement when it sees one. This is a problem you fundamentally
cannot solve by introducing compartmentalizations or meta-levels into
the Seed AI.

I can think of at least two ways to possibly partially-solve this:

(1) Scale back your ambition; rather than worrying about problems such
as P, you could concentrate instead solely upon building a Seed AI
which can solve, and learn to solve, a broad, but not too broad, class
of problems, such as, say the problems in NP. The problem with P is
that there is apparently no way for the person requesting the string
to verify that the answer received is in fact correct…

(2) Forget some of the more extreme ambitions of Seed AI; a Seed AI
alone cannot always solely determine which changes will be useful, but
must rely upon the environment to choose (the Seed AI is borrowing
complexity from the environment to make its decision—this is essential
for problems such as P), i.e., a "hard-takeoff" would be impossible
due to the delay imposed by feedback from the environment.
Furthermore, the Seed AI may be incapable of doing much better than
evolution or random-guessing in some cases.

~Paul Fidika



This archive was generated by hypermail 2.1.5 : Wed Jul 17 2013 - 04:00:49 MDT