Re: [Sheflug] General Ramblings

[Alastair Donlon <nospam [at] netsoc.ucd.ie>]

On Fri, Jun 09, 2000 at 03:23:39PM +0100, Alex Hudson wrote:
> 
> On Fri, 9 Jun 2000, Alastair Donlon wrote:
> 
> > Not anymore. AI's changed a whole lot in the last 10 years. The chess problem
> > has, essentially, been solved 
> 
> 'Chess' certainly hasn't been solved!! Indeed, the theory that chess is
> even *soluable* hasn't been answered.. are you talking more about Deep
> Blue, etc.? That effort was about beating Grandmasters, nothing to do with
> solving chess.
> 

As far as AI is concerned, that's what chess is about. Of course, chess
theory isn't complete. Chess isn't a 'solved' game in the way that tic-tac-toe
is, and never will be. But from an AI perspective, the chess problem is
solved - a computer is able to beat the (arguably) best human player. That
was the challenge set down by John McCarthy et al in 1950 at the conference
that coined the phrase AI.

> > AI research these days has moved away from the classical symbolic problems
> > to the 'real' problems - things like handwriting recognition, speech recognition
> > and robot control.
> 
> That's neural networks.. not something I class as AI, although obviously
> people will argue. Most people don't have a strict definition of AI, but I
> don't include weighted networks, neural networks, heuristics, etc.

Not at all.  You can attempt handwriting recognition with decision trees
or by case-based reasoning or any one of a number of techniques. To me, 
it's not the method which strictly defines whether something is AI or
not, it's the combination of problem and method.  That's why I allow
for NP-hard problems as AI problems - if the technique cuts down the
search space in an intelligent fashion and provides an acceptable
non-optimal solution, then I'm happy to call it AI. 


> 
> > Things like the travelling salesman problem, though, are
> > still _very_ relevant in the commercial world. All timetabling assistants
> > have to work out some non-optimal solution to it, and even the ones on the
> > market on the moment just use some simple heuristics - nowhere near the
> > power that could be provided by modern AI techniques.
> 
> That's even more non-AI ;)) Stochastics & other (essentially) random
> techniques are just about cutting down the search space in some hopeful
> manner, there's not all that much intelligence applied to the situation.

I disagree. The search space is enormous, knowing which swathes to cut
off is an incredibly difficult task. One which even humans have problems
doing. Having said that, NP hard problems are only extensions of the
chess game. But they don't allow for the brute force approach that 
brought about Deep Blue. Quantum computers that can solve NP-hard
problems are as irrelevant as Deep Blue is - they only show that the
problem isn't the best one.

> 
> Maybe it's just me, but I count AI as stuff to do with robots & life
> programs, little else. None of the stuff above counts (IMHO, of course).
> Worse, I've seen papers that try to show Quantum computing to be some
> subset of AI by attempting to apply q-computation to 1-sat, k-sat, etc., 
> problems - seems AI has a very broad definition now. Anything that
> involves NP-hard, complete or other such computationally hard stuff
> (constraint logic), and doesn't have fluffy animals & sweets (life,
> pokemon, etc.) doesn't cut it ;)))


Well that's where we'll have to agree to differ. AI is, of course,
not strictly defined. But I wouldn't restrict it merely to robotics
work or artificial life. I'm not even sure what I could class myself
by your definition - I work on neural network controllers for robotics.
Real robots, that you put a fluffy animal cover on if you like.

A.D.
---------------------------------------------------------------------
Sheffield Linux User's Group - http://www.sheflug.co.uk
To unsubscribe from this list send mail to
- <sheflug-request [at] vuw.ac.nz> - with the word 
 "unsubscribe" in the body of the message. 

  GNU the choice of a complete generation.