BEGIN:VCALENDAR
VERSION:2.0
PRODID:-//talks.bham.ac.uk//v3//EN
BEGIN:VEVENT
CATEGORIES:Optimisation and Numerical Analysis Seminars
SUMMARY:A block-coordinate Gauss-Newton method for nonline
ar least squares - Jaroslav Fowkes (University of
Oxford)
DTSTART:20181122T120000Z
DTEND:20181122T130000Z
UID:TALK3364AT
URL:/talk/index/3364
DESCRIPTION:Nonlinear (nonconvex) least squares problems are u
sed for a range of important scientific applicatio
ns\, such as data assimilation for weather forecas
ting and climate modelling\, where parameter estim
ation is needed in order to specify simulation mod
els that fit observations. In many of these applic
ations\, a model run is computationally expensive
but provides the full vector of simulated observat
ions. However\, calculating the entire Jacobian ma
y be too expensive\, as it may involve additional
model runs along each variable coordinate. We prop
ose a block-coordinate Gauss-Newton method that ca
lculates Jacobians only on a subset of the variabl
es/parameters at a time. We investigate globalisin
g this approach using either a regularization term
or a trust-region model and show global complexit
y results as well as extensive computational resul
ts on CUTEst test problems. Furthermore\, as our a
pproach exhibits very slow rates of convergence on
certain problems\, we design adaptive block size
variants of our methods that can overcome these in
efficiencies.
LOCATION:Nuffield G13
CONTACT:Sergey Sergeev
END:VEVENT
END:VCALENDAR