From newshub.ccs.yorku.ca!ists!helios.physics.utoronto.ca!news-server.ecf!news-server.csri.toronto.edu!rpi!zaphod.mps.ohio-state.edu!qt.cs.utexas.edu!cs.utexas.edu!sun-barr!newstop!sun!amdcad!netcomsv!nagle Mon Dec 16 11:01:16 EST 1991
Article 2053 of comp.ai.philosophy:
Path: newshub.ccs.yorku.ca!ists!helios.physics.utoronto.ca!news-server.ecf!news-server.csri.toronto.edu!rpi!zaphod.mps.ohio-state.edu!qt.cs.utexas.edu!cs.utexas.edu!sun-barr!newstop!sun!amdcad!netcomsv!nagle
>From: nagle@netcom.COM (John Nagle)
Newsgroups: comp.ai.philosophy
Subject: Re: Short List: Things Computers Can't Do
Message-ID: <1991Dec11.213808.20046nagle@netcom.COM>
Date: 11 Dec 91 21:38:08 GMT
References: <1991Dec10.225019.6919@cs.sfu.ca>
Organization: Netcom - Online Communication Services  (408 241-9760 guest)
Lines: 43

shell@cs.sfu.ca (Barry Shell) writes:

>Outright impossible:
>The halting problem
     Only for machines with infinite memory.  For any deterministic
machine with finite memory, a program must either halt or repeat a 
previous state in a finite number of cycles.  If it repeats a previous
state, it is in an infinite loop and will never halt.
>Tiling the plain with 4 coloured 4 sided tiles where like colors meet.

>Solvable but requires human to guess:
>Bin Packing
>Scheduling
      There are pathological scheduling problems for which the optimal
solution is NP-hard.  In many cases, near-optimal solutions can be easily
obtained, as with the travelling salesman problem.  Humans do much worse
than algorithms at these.

>(Above may be classed as solvable but in an impractically long time)
>Therefore they are not practically solvable.

>Solvable in the long term but not very well for now:
>Walking
      6-legged walking is well-understood.  Four-legged walking has
been done a few times.  Two legged-walking and running are hard, but 
see Raibert's "Legged Robots that Balance".  There's been real progress
in the last five years.  The key insight is that once you solve bouncing
up and down in place, running is easy.
>Talking
      Talking is easy.  Listening is hard.
>Vision
      Hard, partly because huge compute power is required.  But as 
power over a few hundred MIPS is applied, the problem is starting to crack.
Check out the CMU NavLab effort.
>Hearing
      Voice recognition is coming along, but still doesn't work very well.
>Learning
      Existing learning systems are still very weak.  Most are optimizers
or statistical classifiers disguised as AI.
>Reasoning with flexibility
      What does this mean?

					John Nagle


