Date: Mon, 20 Feb 2006 01:36:20 -0600 (CST)
Subject: backup strategy and numerical technology
X-UID: 136

The laptop is wiped and everything reinstalled. Most of the toolchains are 
back in place with upgraded versions. Hard disk space is more efficiently 
utilized. So overall, it's better. The overall cost is about two days.

The WiFi kernel/user and webcam kernel device drivers caused the machine 
to become stuck a few times before the kinks were worked out. The last 
straw may have been my remote control application that uses pthreads with 
iostreams. Whether or not there is a kernel oops, when impossible things 
start happening then it is time to minimize the damage and reboot. My 
guess is corruption in the root filesystem.

In this case, the laptop wouldn't boot! I'm not sure why. It could be the 
hard disk has changed. Or it could be the optical drive or the USB 
external hard disk plugged into the laptop. I picked up a 250 GB external 
disk as part of a backup strategy with CDRs. The large external disk is 
only for backups, not for regular use. A reader of this blog once asked me 
what my disaster recovery strategy was. I didn't have one. I think I'm 
lucky as I had a near disaster but not a real one.

I've read enough about Kalman filters to understand that remote driving of 
the robot with live sensor data collection is the first step. The filter 
models the sensors. There's no way to construct the model without 
experimental data. So instead of my hacked remote driving software, 
something more robust is required. Back when battlebots were still a big 
deal, a coworker wondered about the feasibility of a fully digital remote 
control system for robots instead of conventional RC systems (even with 
digital radios). That's basically what I'm writing now.

For simple examples, the Kalman filter leads to highly structured 
numerical linear algebra problems. I've downloaded CLAPACK (C version of 
LAPACK - numerical linear algebra library in FORTRAN) and ATLAS (optimized 
BLAS - basic linear algebra subroutines). ATLAS is work done in the last 
five years. It didn't exist when I was in school. This is encouraging to 
see. Numerical linear algebra technology is not dead.

Even though matrix algebra was my favorite subject, for whatever reason I 
never used COTS (common off the shelf) numerical linear algebra software. 
I always wrote my own. I think part of the reason is the educational 
system. As a student, you must demonstrate knowledge of technology by 
reimplementing it. It's only if you are doing research and development 
that you can use COTS technology.

I have this theory that what Sebastian Thrun (Stanford AI lab director) 
advocates in stochastic, highly numerical robotics and AI is the new 
"subsumption" architecture. People will again build machines that are 
quite astonishing at times but not really understand why they work.