Fundamental Compressionist Philosophy.

[Detlef Morgenstern]

--- Gerry Wolff <gerry at informatics.bangor.ac.uk> wrote: 
> ----- Original Message -----
> From: "Brendan Macmillan" <bren at mail.csse.monash.edu.au>
> To: <gerry at informatics.bangor.ac.uk>
> Cc: "casc mail list" <casc at sanna.com>
> Sent: 26 May 2001 00:25
> Subject: Re: Fundamental Compressionist Philosophy.
> 
> 
> > Knowing what "looks promising" requires prior knowledge.
> > And the decision
> > to use hill climbing requires a hunch that the search space is
> > indeed hilly.
> > Usually, we can present data in a format that reveals its
> > hilliness - but then we ourselves have gone and done the 
> > difficult bit.
> 
> OK, it is possible to call these things 'prior knowledge' and if
> the term is used in that way, you are right.

Which one is 'uphill' from
111000101010:

101000001010     or
111100101010     ??

Why?
Where is 'uphill'?

Are
111000101010
111010101010
near or far neighbours?

Is 
111000101110
between them?

You must know nearly everything about the domain which you want to
search by hill climbing etc.

Altering one bit (not knowing what it 'means') may beam you to the
opposite side of the galaxy.

It must work completely context-free, or it won't work at all.

I see heuristics work only on such global (off-context) parameters as
-- amount of occupied resources
  ("4G units is worse than 2.5G units")
-- training set (rule) reproduction fidelity
  ("failed at 2 samples" is better than "failed at 200 samples")
or the like.

I accept that this is little - compared to the search space volume.

Detlef



__________________________________________________________________
Do You Yahoo!?
Gesendet von Yahoo! Mail - http://mail.yahoo.de