TITLE: Computer Science as Science
AUTHOR: Eugene Wallingford
DATE: February 08, 2007  6:23 PM
DESC: 
-----
BODY:
If you're in computer science, you've heard the derision from
others on campus.  "Any discipline with 'science' in the name
isn't."  Or maybe you've heard, "What we really need on campus
is a department of toaster oven science."  These comments reflect
a deep misunderstanding of what computing is, or what computer
scientists do.  We haven't always done a very good job
<A HREF="http://www.cs.uni.edu/~wallingf/blog/archives/monthly/2005-06.html#e2005-06-22T10_25_42.htm">
   telling our story</A>.
And a big part of what happens in the name of CS really isn't
science -- it's engineering, or (sadly) technical training.
But does that mean that no part of computer science is 'science'?

Not at all.  Computer science is grounded in a deep sense of
empiricism, where the scientific method plays an essential role
in the process of doing computer science.  It's just that the
entities or systems that we study don't always spring from the
"natural world" without man's intervention.  But they are
complex systems that we don't understand thoroughly yet --
systems created by technological processes and by social
processes.

I mentioned an example of this sort of research from my own
backyard back in
<A HREF="http://www.cs.uni.edu/~wallingf/blog/archives/monthly/2004-12.html#e2004-12-30T08_06_20.htm">
   December 2004</A>.
A master's student of mine, Nate Labelle, studied relationships
among open-source software packages.  He was interested in seeing
to what extent the network of 'imports' relationships bore any
resemblance to the social and physical networks that had been
studies deeply by folks such as
<A HREF="http://www.nd.edu/~alb/">
   Barabasi</A>
and
<A HREF="http://smallworld.columbia.edu/watts.html">
   Watts</A>.
Nate conducted empirical research: he mapped the network of
relationships in several different flavors of Linux as well
as a few smaller software packages, and then determined the
mathematical properties of the networks.  He presented this
work in a few places, including a paper at the
<A HREF="http://necsi.org/community/wiki/index.php/ICCS06/226">
   6th International Conference on Complex Systems</A>.
He concluded that:

<BLOCKQUOTE><EM>
... despite diversity in development groups and computer system
architecture, resource coupling at the inter-package level creates
small-world and scale-free networks with a giant component containing
recurring motifs; which makes package networks similar to other
natural and engineered systems.
</EM></BLOCKQUOTE>

This scientific result has implications for how we structure
package repositories, how we increase software robustness and
security, and perhaps how we guide the software engineering
process.

Studying such large systems is of great value.  Humans have
begun to build remarkably complex systems that we currently
understand only surface deep, if at all.  That's is often a
surprise to non-computer scientists and non-technology people:
The ability to <EM>build</EM> a (useful) system does
<STRONG>not</STRONG> imply that we <EM>understand</EM> the
system.  Indeed, it is relatively easy for me to build systems
whose gross-level behavior I don't yet understand well (using
a neural network or genetic algorithm) or to build systems
that perform a task so much better than I that it seems to
operate on a completely different level than I do.  A lot of
chess and other game-playing programs can easily beat the people
who wrote them!

We can also apply this scientific method to study processes that
occur naturally in the world using a computer.  Computational
science is, at its base, a blend of computer science and another
scientific domain.  The best computational scientists are able
to "think computationally" in a way that qualitatively changes
their domain science.

As Bertrand Russell wrote a century ago, science is about
<EM>description</EM>, not logic or causation or any other naive
notion we have about necessity.  Scientists describe things.
This being the case, computer science is in many ways the
ultimate scientific endeavor -- or at least a foundational one
-- because computer science is <EM>the science of description</EM>.
In computing we learn how to describe thing and process better,
more accurately and more usefully.  Some of our findings have
been surprising, like the
<A HREF="http://www.cs.uni.edu/~wallingf/blog/archives/monthly/2007-02.html#e2007-02-06T22_31_26.htm">
   unity of data and program</A>.
We've learned that process descriptions whose behavior we can
observe will teach us more than static descriptions of the
same processes left to the limited interpretative powers of
the human mind.  The study of how to write descriptions --
programs -- has taught us more about language and expressiveness
and complexity than our previous mathematics could ever have
taught us.  And we've only begun to scratch the surface.

For those of you who are still skeptical, I can recommend a
couple of books.  Actually, I recommend them to any computer
scientist who would like to reach a deeper understanding of
this idea.  The first is
<A HREF="http://www.psy.cmu.edu/psy/faculty/hsimon/hsimon.html">
   Herb Simon</A>'s
seminal book
<A HREF="http://www-static.cc.gatech.edu/~jimmyd/summaries/simon1969-a.html">
   The Sciences of the Artificial</A>,
which explains why the term "science of the artificial" isn't
an oxymoron -- and why thinking it is is a misconception about
how science works.  The second is
<A HREF="http://eksl.cs.umass.edu/~cohen/">
   Paul Cohen</A>'s
methodological text
<A HREF="http://eksl-www.cs.umass.edu/emai.html">
   Empirical Methods for Computer Science</A>,
which both teaches computer scientists -- especially AI students
-- how to do empirical research in computer science.  Along the
way, he demonstrates the use of these techniques on real CS
problems.  I seem to recall examples from machine learning, which
is perticularly empirical in its approach.
-----