Algorithmic Differentiation for Adjoint-based Optimization

Decanato - Facoltà di scienze informatiche

Data d'inizio: 28 Settembre 2010

Data di fine: 29 Settembre 2010

DATE: Tuesday, September 28th 2010
PLACE: USI Università della Svizzera italiana, room CC250, Main building (Via G. Buffi 13)
TIME: 13.30

ABSTRACT:
In this talk we discuss the exact computation of derivatives using the technique of Algorithmic Differentiation (AD).
Complexity estimates for the two basic approaches, i.e.
the forward mode and the reverse mode, will be given.
This includes for example the "cheap gradient result", i.e.
that computational complexity for evaluating the gradient of a scalar-valued function can be bounded above by a computational complexity of the function evaluation multiplied by a small constant.
This upper bound allows the usage of exact gradient information for numerous optimization problems.
The exact derivative information in combination with structure exploitation allows the optimization of rather complicated processes and structures. Several examples will be shown in this talk including applications from aerodynamics and nuclear optics.

BIO:
2009 - Professor, University of Paderborn
2008 Habilitation, Technical University of Dresden
2007 Substitute of the Chair "Mathematics I", University of Regensburg
2007-2008 Assistant Professor of "Analysis and optimization of computer models", Technical University of Dresden
2003-2006 Head of the research group "Analysis and optimization of computer models", Technical University of Dresden
1999 PhD in Mathematics, Technical University of Dresden
1997-1999 research assistant in the DFG research group "Identification and optimization of complex models based on analytical sensitivity calculations", Technical University of Dresden
1996 Diploma in Industrial Mathematics, University of Bayreuth
1991-1996 Studies (Business Mathematics), University of Bayreuth

 

HOST: Prof. Rolf Krause