IEEE - Aerospace and Electronic Systems - October 2022 - 46

Interview With Fred Daum
Fred in action, pointing out the often-unmentioned problem ofdegeneracy in particle filtering (a common pet peeve).
BMEWS solid state upgrade, relocatable over-the-horizon
radar (ROTHR), SPY-3, JLENS, UEWR, THAAD
radar, AN/TPY-2, SBX, and many others. I designed,
analyzed, and tested essentially all of the real-time algorithms
for long-range phased array radars built by the
USA in the last four decades. These algorithms include
extended Kalman filters, batch least squares filters, and
data association algorithms [e.g., multiple hypothesis
tracking (MHT) and NNJPDA] but are designed to
account for unresolved sensor measurements, real-time
radar waveform scheduling, Bayesian object classification,
discrimination of missiles from satellites, and other
stuff, fancy detection algorithms, calibration of tropospheric
and ionospheric errors, calibration of diverse
monopulse effects caused by the environment, and more.
But I was also called upon to fix sonar, GPS, and other
systems. One of the most challenging tasks was to design
an algorithm to decide whether a given track is a ballistic
missile (ICBM or SLBM) rather than something else: satellite,
ionospheric clutter, jamming, chaff, aircraft, noise,
ducted clutter, birds, unresolved measurements, data
association errors or maybe a broken, or miscalibrated
radar. The customer wanted an extremely low probability
of decision errors, because such errors might start World
War III; one USAF General told me this very explicitly
to make sure that I understood how important this algorithm
was. Making the problem more difficult was the
fact that we had to make such decisions about once per
second. The customer wanted an error less than once in a
blue moon. This implies an error probability of about
one in a zillion, which would be about 5 or 10 sigmas
assuming a zero mean univariate Gaussian density. But
of course, the errors are not Gaussian, not zero mean,
and not univariate. The errors do not correspond to any
46
named textbook probability density, and it is certainly
not stationary. In any case, we started with a carefully
designed Bayesian decision rule and proceeded to make
it work in the real world after about 3 months of experience,
ad hoc fixes, tuning of parameters, and a prefrontal
lobotomy to our extended Kalman filter, somewhat like
dropout in modern deep learning. Textbook noise never
caused decision errors, because our Bayesian algorithm
was designed to handle this case, assuming that the error
covariance matrices going into it were reasonably good.
What caused the most errors were ducted clutter and ionospheric
clutter and false cross-range observability in our
EKF, along with occasional undiagnosed radar hardware
glitches. We got the algorithm to work about 10 times better
than required, on schedule and within budget, to the
delight ofall concerned. Another similar story with a happy
ending concerned tracking missiles in the boost phase with
a very large Doppler shift due to the rocket exhaust plume
interacting with the amazingly large range-Doppler coupling
of our transmitted radar waveforms. Accurate tracking
despite highly uncertain missile thrust acceleration
profiles and staging and an extremely dense multiple object
environment was a cakewalk by comparison. There are
many other stories about getting such radar systems to
work in the real world. I loved it. I learned a lot of new
physics about the ionosphere and troposphere and plasmas
(e.g., rocket exhaust plume) and the thermal expansion of
the aluminum face of our beautiful phased array antennas.
But Yaakov probably has more questions for me, so we
will leave it there for now.
In addition, I loved doing research on fancy nonlinear
filters and giving technical talks for the IEEE and SPIE as
well as giving invited talks at schools, such as MIT, Harvard,
Yale, CalTech, the Technion, Ecole Normale
IEEE A&E SYSTEMS MAGAZINE
OCTOBER 2022

IEEE - Aerospace and Electronic Systems - October 2022

Table of Contents for the Digital Edition of IEEE - Aerospace and Electronic Systems - October 2022