A Survey of Numerical Algorithms for
Optimization
Kaj Madsen, Department of Mathematical Modelling, Technical University of
Denmark
Monday June 19, 2000, 2:00pm
CRL, Room B102
Abstract:
During the past 40 years an important part of
numerical analysis has been the development of methods for minimizing
nonlinear functions. A large and elegant theory
supporting the methods has been worked out.
Most of the methods are
iterative, and they are based on Taylor's formula: At the current
iterate x a model of the nonlinear function f is built, based on function
and gradient information available at x. This model is valid at some
neighborhood of x and forms the basis for finding the next iterate.
The most successful class of methods of this type is the Quasi-Newton
methods which can now be considered as fully developed.
If the function f has special characteristics (e.g.,
being a sum of squares or a minimax function)
local models have been used which simulate the
special structure of f.
In global optimization other types of methods have
been more common. Most often they are based on stochastics. Another prominent
example of a global optimization strategy based on more general information
about f is the interval method. This method even has a built-in
result verification. During the years many
researchers have been trying to build local models based
on information from several points rather than using the Taylor philosophy.
So far this has not been generally successful as compared with the
Quasi-Newton principle. However, recent developments have shown promising
indications in this direction, mostly through strategies devoted to
special applications. Examples are surrogate modeling in aircraft design
and space mapping techniques for electromagnetic optimization.
Refreshments and snacks will be served.
For a biography of Kaj Madsen please visit his website
http://www.imm.dtu.dk/documents/users/km/homepage.html
For updated information please visit http://www.sos.mcmaster.ca/upcomingevents.htm |