Re: The Demise of Computationalism?

Michael Olea wrote:
Neil W Rickert wrote:

Pei Wang puts it like this:

"Abstract: This paper is about the philosophical and methodological
foundation of artificial intelligence (AI). After discussing what is a good
"working definition", "intelligence" is defined as "the ability for an
information processing system to adapt to its environment with insufficient
knowledge and resources''. Applying the definition to a reasoning system,
we get the major components of Non-Axiomatic Reasoning System (NARS), which
is a symbolic logic implemented in a computer system, and has many
interesting properties that closely related to intelligence. The definition
also clarifies the difference and relationship between AI and other
disciplines, such as computer science. Finally, the definition is compared
with other popular definitions of intelligence, and its advantages are
argued."

http://www.cogsci.indiana.edu/farg/peiwang/papers.html#intelligence
It's an open question. Your comment has no bearing on that question.
The open question is whether a computer can generate a similar range
of intelligent behavior to that produced by people.


We mentioned Nars a little over a year ago in "Material vs. Immaterial"
including this quote, for the benefit of other readers,

"In the following we will introduce an AI system
called NARS (Non-Axiomatic Reasoning System).
It is a reasoning system, in the sense that it has
its language, semantics, inference rules, memory,
and control mechanism. However, it is not axiomatic,
algorithmic, or formal, in the most natural senses
of these notions." ...

"The Logic of Intelligence" by Pei Wang continues with:

...."However, as a virtual machine, NARS can be based on
another virtual machine which is a pure-axiomatic system,
as shown by its implementation practice, and this fact
does not make the system "axiomatic". If we take the
system's complete experience and response as input and
output, then NARS is still a Turing Machine that definitely
maps inputs to outputs in finite steps. What happens here
has been pointed out by Hofstadter as "something can be
computational at one level, but not at another level"
[Hofstadter, 1985], and by Kugel as "cognitive processes
that, although they involve more than computing, can still
be modelled on the machines we call `computers'" [Kugel,
1986]. On the contrary, traditional computer systems are
Turing Machines either globally (from experience to response)
or locally (from question to answer)."

SH: This reminds me of Marcus Hutter and Solomonoff Induction
"My research at IDSIA was centered around the AIXI model,
which is a mathematical top-down approach to AI, related
to Kolmogorov complexity, algorithmic probability, universal
Solomonoff induction, Occam's razor, Levin search, sequential
decision theory, dynamic programming, reinforcement learning,
and rational agents."

http://www.hutter1.net/ai/index.htm

Abstract: "Decision theory formally solves the problem of
rational agents in uncertain worlds if the true environmental
probability distribution is known. Solomonoff's theory of
universal induction formally solves the problem of sequence
prediction for unknown distribution. We unify both theories
and give strong arguments that the resulting universal AIXI
model behaves optimally in any computable environment. The
major drawback of the AIXI model is that it is uncomputable.
To overcome this problem, we construct a modified algorithm
AIXI^tl, which is still superior to any other time t and
space l bounded agent. The computation time of AIXI^tl is
of the order t x 2^l."

SH: This is background to introduce some material from a
paper of Solomonoff Induction,

"Now that we have the basic rudiments of measure theory we
now consider how this applies in our context of strings and
sequences. We will soon be interested in predicting digits
in a sequence after having seen a finite number of initial
digits. This means that we need to have probability measures
defined over sets of sequences which have common initial digits."
....
"Solomonoff's induction method is an attempt to design a
general all purpose inductive inference system. Ideally such
a system would be able to accurately learn any meaningful
hypothesis from a bare minimum of appropriately formatted
information. Before trying to define such an inference system
and analyse its behaviour, we first need to form a reasonable
idea as to what such a system might look like. To help us do
this, let's imagine some sort of super intelligent device or
being that operates as a perfect inductive inference system.
For the sake of our thought exercise we will call this machine
or being Zed.
Our question is: What properties will Zed have? Firstly, Zed
mustn't be too narrow minded as to what could potentially be a
correct hypothesis. It is however clear that any non-computable hypothesis would not be of much use to anybody. Hence restricting
Zed's set of possible hypotheses to only the computable ones seems reasonable enough. Exactly what this means will become clear when
we formalise all these ideas later on. Next consider what sort of knowledge Zed has about things before processing any data.

For Zed to be truly all purpose, Zed must be able to function in
situations where no prior information about the system under
investigation is avaliable. For example, in a totally artificial
inductive inference problem, even complete knowledge of all the
laws of physics would be of no help. This is not to say that Zed shouldn't be able to utilise prior information, but simply that
prior information is an extra rather than an essential part of
Zed's operation. Thus if we consider Zed's initial state to be
independent of the problem at hand, then it follows that this
state must be one of complete ignorance about the nature of the
system under investigation.

Now that we have some idea about Zed's initial state and the set
of potential hypotheses that Zed is going to consider, we next
look at what actions Zed will need to be able to perform.
Obviously Zed will need to be able to process information in
order to determine which hypotheses are likely and which are
unlikely or even impossible. Perhaps Zed's first source of such
information would be the prior information mentioned above, that
is; any knowledge relevant to the system under investigation that
comes from work conducted prior to the current investigation. For
example, knowledge of the laws of physics would often be helpful
when studying real physical systems. It might be the case that
others have studied similar systems before. It could even be the
case that the correct hypothesis is already known! Thus, just as
any serious scientist checks what information already exists
before contributing his own ideas and performing his own
experiments etc, Zed must at least have the ability to utilise
such information when available."

SH: This ability to access prior information seems like the
same functionality that CyC aims to provide. Perhaps instead
of the current 1.6 million inter-connected assertions, 1.6
billion would provide sufficient efficiency. But this is a
matter of more or less of the same kind of functionality, not
an example of different types of functionality.
....
"Having exhausted all information deriving from pervious work,
the next avenue must be for Zed to gather further information
himself through experimentation and observation. This information
will further refine Zed's estimate of the true hypothesis. It is
then possible that the experimental information might point to
further areas of previous work that should be taken into account.
Perhaps it will indicate new sets of experiments that should be
carried out. In this way Zed gathers more and more information,
continually updating and re¯ning his degree of belief in the
various possible hypotheses. This process is of course the process
by which all scientists operate. Various hypotheses gain or lose
favour in the light of new information or even old information
which has been overlooked. Bayes' theorem provides us with such
a method, however it demands that we assign prior probabilities to
each hypothesis. This is where the real innovation in Solomonoff's technique lies and we will examine his solution and its
consequences in the following section."

I'm not sure it is dynamic however,
Stephen
.

Relevant Pages

  • Re: Cantor Confusion
    ... the notion of a sequence derives really from an inductive definition ... Since ZFC has no trouble modeling the natural ... But ZFC does have considerable difficulty dealing with infinite ... proof by induction that these "infinite quantities" were also simply ...
    (sci.math)
  • Re: Uncountably of Super Infinite Sets
    ... then A union B <= C union D. ... Do you PROVE Peano induction? ... then it is true of every element of the sequence. ... but inductive proof doesn't introduce a new term into the ...
    (sci.math)
  • Re: [Discrete Math] Divisibility within subsequences of the Fibonacci Sequence
    ... denotes the kth Fibonacci Number, ... Also, again from prior posts, the topic is "induction", and the goal, ... sequence, but rather, to provide an exercise for the student to apply ... The first result solves most first level problems ...
    (sci.math)
  • Re: product of n successive positive integers
    ... had given an inductive one and wanted a direct proof. ... sequence as a McLaurin sequence with f=^k? ... were to prove that one, you'd go about it by induction, wouldn't you? ... That's why I named McLaurin sequence: ...
    (sci.crypt)
  • Re: approaching a proof
    ... Mitch Harris wrote: ... >> sequence of 'proofs' in that system that APPROACH a proof of the ... > induction (i.e. choose your system to be arithmetic without ... What axioms and rules are you assuming? ...
    (sci.logic)