Re:What I wish systems researchers would work on
by graycat on Thursday May 16, 2013 @10:09AM
(#5718669)
For where computing will be based, currently
there are big cultural issues:
With one
paper I published, in a computer science
journal for a problem in computer science
but basically some applied math, I had to discover that, really, it was tough for the computer science
community to review the paper. I sent an
'informal' submission to a nice list of the top
journals appropriate for the paper. From one
Editor in Chief of one of the best journals and
a chaired professor of computer science at one
of the best universities I got back "Neither I
nor anyone on my board of editors has the
mathematical background to review your paper."
Then I got a similar statement from another
such person. For one at MIT, I wrote him
background tutorials for two weeks before he
gave up. Finally I found a journal that
really wanted the paper, but apparently
in the end only the editor in chief reviewed
the paper and did so likely by walking it
around his campus to the math department
to check the math and then to the computer
science department to check the relevance
for computer science.
I had to conclude: For the good future of
computer science via math, the current crop of
CS profs just didn't take the right courses
in college and graduate school.
That's one side of the cultural divide. For
the other side, (1) the math departments
always want to be as pure as possible and
f'get about applied math and (2) in applied
math really don't like doing computer science.
Some obscure math of relativity theory,
maybe, but mostly nothing as practical as
computer science.
Of course, the big 'cross cutting' exception
is the problem P versus NP, apparently first
discovered in operations research (integer
linear programming, yes, in NP-complete),
later in computer science with SAT, and now
at times taken seriously in math, e.g., at
Clay Math.
Here's a 'reason' for the math: When we
write software, we need something prior to
the software as, say, a 'specification', that
is, saying what the software is to do.
Now, where is computer science going to get
that specification? A big source has been
just to program what we know how to do
at least in principle just manually. After
that, computer science starts to lose it
and drift into 'expert systems' (program,
with 'rules' and Forgy's RETE algorithm
what an expert says), intuitive heuristics,
and various kinds of brute-force fitting,
machine learning, neural networks, where
basically where we fit to the 'training'
data, test the fit with the rest of the
data, and stop when get a good fit. So
we throw fitting methods at the data
until something appears to stick.
We need more powerful means of getting that
prior specification. The advantage of
math is that it can start with a real problem,
formulate it as a math problem, solve the
math problem, and then let the math solution
and what is says we needed to do in manipulating
the data be the specification for the core
software. E.g., if want to design an airplane
on a computer, then start with the applied
math of structural engineering, mechanical
engineering, and aeronautical engineering, program
that applied math, and then design the plane.
For software to navigate a spacecraft to the
outer planets, start with Newton's second
law and law of gravity, get the differential
equations of motion, get some numerical
means of solution, and then program what the
numerical analysis says to do.
We need things solid and prior to the software
to know what software to write, and basically
that is we need a math solution first.
For computing itself, as in monitoring, we
can call that a problem in statistics --
more applied math.
A lot in computer load balancing is
some serious applied math. Actually
optimal job scheduling is awash in
NP-complete optimization problems.
Or, we used to have 'metaphysics'. Then
physics became mathematical and made
real progress. Basically the solid logical
chain of correctness given by math theorems
and proofs is just too darned hard to
compete with or, thus, ignore.
With one paper I published, in a computer science journal for a problem in computer science but basically some applied math, I had to discover that, really, it was tough for the computer science community to review the paper. I sent an 'informal' submission to a nice list of the top journals appropriate for the paper. From one Editor in Chief of one of the best journals and a chaired professor of computer science at one of the best universities I got back "Neither I nor anyone on my board of editors has the mathematical background to review your paper." Then I got a similar statement from another such person. For one at MIT, I wrote him background tutorials for two weeks before he gave up. Finally I found a journal that really wanted the paper, but apparently in the end only the editor in chief reviewed the paper and did so likely by walking it around his campus to the math department to check the math and then to the computer science department to check the relevance for computer science.
I had to conclude: For the good future of computer science via math, the current crop of CS profs just didn't take the right courses in college and graduate school.
That's one side of the cultural divide. For the other side, (1) the math departments always want to be as pure as possible and f'get about applied math and (2) in applied math really don't like doing computer science. Some obscure math of relativity theory, maybe, but mostly nothing as practical as computer science.
Of course, the big 'cross cutting' exception is the problem P versus NP, apparently first discovered in operations research (integer linear programming, yes, in NP-complete), later in computer science with SAT, and now at times taken seriously in math, e.g., at Clay Math.
Here's a 'reason' for the math: When we write software, we need something prior to the software as, say, a 'specification', that is, saying what the software is to do. Now, where is computer science going to get that specification? A big source has been just to program what we know how to do at least in principle just manually. After that, computer science starts to lose it and drift into 'expert systems' (program, with 'rules' and Forgy's RETE algorithm what an expert says), intuitive heuristics, and various kinds of brute-force fitting, machine learning, neural networks, where basically where we fit to the 'training' data, test the fit with the rest of the data, and stop when get a good fit. So we throw fitting methods at the data until something appears to stick.
We need more powerful means of getting that prior specification. The advantage of math is that it can start with a real problem, formulate it as a math problem, solve the math problem, and then let the math solution and what is says we needed to do in manipulating the data be the specification for the core software. E.g., if want to design an airplane on a computer, then start with the applied math of structural engineering, mechanical engineering, and aeronautical engineering, program that applied math, and then design the plane. For software to navigate a spacecraft to the outer planets, start with Newton's second law and law of gravity, get the differential equations of motion, get some numerical means of solution, and then program what the numerical analysis says to do.
We need things solid and prior to the software to know what software to write, and basically that is we need a math solution first.
For computing itself, as in monitoring, we can call that a problem in statistics -- more applied math.
A lot in computer load balancing is some serious applied math. Actually optimal job scheduling is awash in NP-complete optimization problems.
Or, we used to have 'metaphysics'. Then physics became mathematical and made real progress. Basically the solid logical chain of correctness given by math theorems and proofs is just too darned hard to compete with or, thus, ignore.