Understanding descriptions
- From: heiska mikko <mikk2lnx@xxxxxxxxxxx>
- Date: 18 Apr 2007 17:13:39 -0700
State class descriptions are a neglected _part_ of A(G)I development
(including reinforcement learning type approaches to A(G)I
development)
(
Some of you may feel it improbable that this is really about general
intelligence(in addition to other, narrow kind of AIs)(or, to be more
precise, general AI's seed AI), but is it somehow known that the
workings of general intelligence must feel intuitive and seem
obviously right when described?
(
Partly conversely, part of AGI related philosophy may be described as
school textbook ideas, but the extreme difficulty is in knowing which
such ideas to include in programming to describe components.
Speaking of decisions about component inclusion, AI development must
have philosophical test(s) / thought experiment(s) for evaluating
theorized/ thought software component's potential for componentness in
the input-output-relation causing system called "general
intelligence". That is because any such component probably won't exist
at the time that it's need is conjectured for the first time (it may
exist by coincidence or because part of narrow-AI development was on
right track) and it is not efficient to do code from every whim of
conjecture, and because practical tests on partial systems would not
reveal enough.
It is a confusing parallel between AI-field's substance and AI-field's
metalevel, that AI itself must have a component for evaluating
speculated software- or physical component's or phenomenon's potential
for componentness in some effect causing system, or cause-effect- or
data-input-output-relation causing system or data-output causing
system(fixed-file-decompressing/generating program), for it's
day-to-day work.
Final decision about componentness is made by considering and
calculating the component candidate as a part of whole plan.
))
The main reason (or, the most interesting explanation) why chess AIs
have used enormous amounts of computation but still barely win best
humans, is that they have'nt processed board state class descriptions,
just evaluated millions of individual board states by working in the
lowest possible level of abstraction / level of description.
If chess AI could handle board state descriptions, that instead of
meaning just 1 position, could refer in one little statement set to a
class of positions that is meaning thousands or millions of positions,
then it would be possible that a software running in a mere mobile
phone could win anybody.
Similarly, even AI could not program ordinary lenght software from
algorithm invariant specification by adding more lines after another
on the basis of previous lines and by rearranging existing
lines(making it less partial and less wrong). That would be practical
for only short programs. Similar thing applies for bridge plans and
natural language sentences. AI must find/produce descriptions of
classes, software component classes, bridge part classes, engine part
classes, word or sentence sequence meaning classes etc. That makes the
goal setting statements less vague, while the procedures of 'less
partial' and 'less wrong' get also used in their proper turn. Only at
the last and lowest description level AI settles to particular
instances and - if applicable - points in continuum.
Ever more efficient search methods by limiting search are not enough
and lack some ways of problem solving that can be described as having
philosophical-level counterparts in nature.
Depending on the nature of the problem and point of view to the
problem, we might call the statement set that means and sets AI's
motivation, as "desired state", "goal", "optimization target",
"specification", "end state", "constraint", "result"... Now 'goal' is
the choice.
Broadest possible meaning of 'goal' and it's minimal size definition
should be used, as with the concept of 'symbol', for example. All
version-specific and context-specific meanings should be left out, if
it is possible to decide what those are.
A 'goal' may not be just something that AI's design(AI's output that
is blueprint, source code or other kind of design sort of thingie)
must do to matter and energy(as in mechanical and civil engineering),
or to data(as in programming). Piece of data meaning 'goal' and an
output resulting from it (+from previous learning results), could also
be something that we aren't normally aware of and which is difficult
to refer at, but without which the AI could not be considered
"general".
(This usenet message may require extensive parsing, but if you wan't
something easy, don't do AI.)
Mik-ko Heis-ka
.
- Follow-Ups:
- Re: Understanding descriptions
- From: Vend
- Re: Understanding descriptions
- From: Stephen Harris
- Re: Understanding descriptions
- Prev by Date: Re: consciousness
- Next by Date: help test a web service: send pictures to mobile phones
- Previous by thread: Making a Turing Tape with "quantum" interactions between head and cells
- Next by thread: Re: Understanding descriptions
- Index(es):
Relevant Pages
|
Loading
