I sure don't update this blog very often. So, let's see...
Since I've posted last:
John has replaced fear.incarnate.net, our long-standing webserver,
with a new box. This one has 750gig mirrored disks, whereas the old
one was constantly running out of space. Of course, since they're
John's disks, one of them died already. (Vendetta's DB server's hot
spare disk was DOA, and fear's two mirrored disks have died at the
same time in the past, and then there was the Micropolis disk
debacle.. in short, John has the worst luck with hard disks)
Apparently the warranty on them is good until 2012, so that's good.
Nanoblogger has been upgraded, and as a result my site got a little
facelift. Woo. It's.. uglier, I think.
I found out that my wife's coworker Jen's husband
Damon works at CarSpot.com,
where
Nick Purvis worked until
recently. I think Nick was a cofounder. Jen and Damon live right
down the street from us; the first time I met them was at their
wedding. Small world, I guess.
I also want to point out, belatedly, that I made it to 26th on the
ICFP 2007
contest, and with Ray's help our combined Guild Software effort
made it to 22nd. I didn't get to spend much time at all on this
contest as I was busy that weekend, but I'm pleased with how well we
did given the effort. (A sort of post-mortem is
here).
Speaking of programming contests, Damon's site reminded me of
Project Euler. It's pretty fun,
highly recommended. (see
my
profile - but you must be logged in for that).
I was also in the ~200s ranking (nothing to write home about) on the
Netflix Prize back in March but
I've fallen even lower (607 as of this writing - now that's
embarassing) since then. I used a maximum-likelihood singular value
decomposition approach except with a conjugate gradient optimizer
(L-BFGS, to be specific, a quasi-Newton method), which is a hell of a
lot more efficient than any gradient descent method with learning
rates that a lot of other people are using. But I didn't devote all
that much CPU time to it, and I need to do a fair amount of
preprocessing of the data to be more competitive. Seems a lot of
people are getting a handle on this very interesting problem.
The Netflix challenge been very educational for me as I've been able
to put a lot of stuff I learned from
David
MacKay's excellent book into context. If I come back to the
problem I'll probably try Hamiltonian Monte Carlo or Gibbs sampling in
order to get more than one sample to marginalize over.
Also, watch for the IOCCC 2006 results announcement in November...
If one of the two people who actually read this blog would like
elaboration on any topic I would be happy to oblige.