Argonne National Laboratory

MCS researchers coedit proceedings on automatic differentiation

July 22, 2013

The fundamental goal behind differentiating numerical computations is to be able to generate — ideally automatically — efficient derivative code for models implemented as computer programs. Achieving this goal in practice has proved challenging, especially for numerical models on parallel architectures.

To track the progress in addressing the challenges, and to present the state of the art in automatic differentiation (AD), every four years the scientific community holds a conference focusing on AD applications in science and engineering, its theory, and the development of compiler-based tools and web-based differentiation services. The Sixth International Conference on Automatic Differentiation (AD2012) held in Fort Collins, Colorado, last summer, continued this quadrennial conference series.

In the preface to the book, the editors explain that they adopted the word algorithmic in the title to “better reflect the reality of AD usage and the research results presented in the book.”

The book comprises 31 chapters, including new implementations in graphical modeling environments, new developments in AD algorithms targeting areas such as nonsmooth functions, and applications ranging from the Earth sciences to chemistry and nuclear engineering. The variety of topics should help both AD and non-AD experts gain insight into AD capabilities and approaches for improving efficiency when computing derivatives.

S. Forth, P. Hovland, E. Phipps, J. Utke, and A. Walther, eds., Recent Advances in Algorithmic Differentiation, Lecture Notes in Computational Science and Engineering, vol. 87, Springer, 2012.