Group for Research in Decision Analysis

Algorithmic Differentiation (AD) in Python

Sebastian F. Walter Humboldt-Universität zu Berlin, Germany

In the talk we discuss what kind of derivatives can be evaluated efficiently using AD (both in theory and practice) and give a brief overview of univariate Taylor polynomial arithmetic. We also describe how numerical linear algebra functions like the QR decomposition fit into the framework and show how mixed partial derivatives can be evaluated using polarization identities. We show examples and give pointers to available software in Python.