Hi,
About execution time, in fact we do restrict execution time; when doing automated program learning we currently use a very crude method and set an execution-time cutoff and eliminate from consideration all programs that don't produce an answer before the cutoff is expired.
Of course, I don't expect a brief exchange of emails to convince you or anyone that we've solved any important problems. We're writing up a paper describing our program learning methodology in some detail; but that won't necessarily convince you our method will work as generally as we believe, however... it will just consider some relatively simple test cases...
-- Ben
Yan King Yin <[EMAIL PROTECTED]> wrote:
> > I want to ask: is your class of algorithms guaranteed to terminate
> > in a *bounded* time? If there is no such guarantee then things may get
> > very complicated, bordering on the undecidable.
>
> No guarantees -- merely "probably approximately correct."
>
> That's the way intelligence is, IMO
PAC-learning is generally insufficient to solve the learning problem,
and some restrictions are usually needed. Maybe restricting on the
time complexity is a good idea, or using minimum description length.
> > OTOH, if it is time bounded (perhaps it contains no loops, no recursion
> > or specific kinds of recursion only, etc),
>
> It's completely general, but is probabilistically biased toward simplicity,
> e.g. it's capable of general recursion but is strongly biased toward some
> simple forms of recursion.
If "simple" means some program length measure, then you may get into
trouble because even very simple programs may be undecidable. Why not
restrict on execution time as well? But if you insist on wandering in
algorithmic space...
> > The problem is when we have already processed the input space, then
> > what does the resulting representation look like? Does it contain a
> > relatively large number of concepts (like millions), or is it highly
> > structured with few concepts on each level? We don't know now, but my
> > model is assuming the first case. If it is the second case then your
> > methods may be better. I suspect your approach is more suitable for
> > things like theorem proving or other specific-domain problems...
>
> Well, we are explicitly trying to create a general intelligence, not a
> domain-specific "narrow AI" program.
>
> We can deal with millions or billions etc. of concepts. For some purposes
> we embed concepts in an n-dimensional space and use n-dimensional metric
> structure and topology, effectively treating the space of concepts as a
> continuum. This is useful for example if you want to "mutate" a concept
> into a similar one.
Well, you have not convincingly demonstrated how your model can do
that efficiently, especially when you take no measures to restrict
execution time. I'm waiting to see a more detailed description of it.
YKY
--
_______________________________________________
Find what you are looking for with the Lycos Yellow Pages
http://r.lycos.com/r/yp_emailfooter/http://yellowpages.lycos.com/default.asp?SRC="">
-------
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]
