On Wed, May 21, 2008 at 8:34 AM, Eleanor McHugh
<eleanor / games-with-brains.com> wrote:
> On 21 May 2008, at 13:40, Albert Schlef wrote:
>>
>> Eleanor McHugh wrote:
>>>
>>> [...]
>>> I started out in physics and I've spent a fair amount of my
>>> professional career arguing with very capable CS experts over all
>>> kinds of problems which are elementary with my background but have
>>> them scuttling back to graph theory
>>
>> Could you elaborte here? It sounds interesting.
>
> Not easily without getting sued for breaching various NDAs ;)
>
> Generally though there's a deep-seated belief in CS that because computers
> deal in defined states formal abstractions such as graph theory are the best
> way to tackle a number of often-intractable optimisation problems. For
> simple systems that's true, but as your state-space explodes you end up with
> solutions which are easily expressed elegantly in Lisp or whatever but which
> in practice are completely useless due to their processing requirements.
>
> Anyone who comes from a science or engineering background will recognise
> this as a granularity issue: a state-space has a level of granularity
> beneath which additional precision in differentiating data points becomes
> completely irrelevant in determining real-world behaviour. The classic
> example is the disconnect between classical and quantum mechanics, but it
> occurs in all experimental fields.
>
> Unfortunately many very good CS people still have a mathematician's bias
> towards a perfect answer rather than a working approximation and look to
> maths for their abstractions rather than seeking them in the real world and
> instead getting the granularity of their systems right. In my personal
> experience this then leads to long and involved discussions which go round
> and round in circles for months on end until projects get cancelled because
> the perfect answer is too costly to deploy, and the working approximation
> isn't provable for all cases.
>
>
> Ellie

Albert, a good example of what Ellie is saying would be -- in the ME
world -- like approximating a complex vibration with small
oscillations of a spring (doesn't have to be a spring; i.e., it could
be a very large 3-d pendulum with small movements).  Once you do that,
the real-world problem becomes incredibly easier than factoring in
many parameters not required.  In effect, you solve the problem
ideologically before you brute-force attack it with a turing machine.
Granularity.

Todd