Adjoint Compiler Technology and Standards (ACTS)

Automatic differentiation is an automatic, compiler-like software tool that can read in a user’s subroutine, analyze it and then automatically write a new subroutine that will evaluate the partial derivatives of the functions implemented by the subroutine with respect to the subroutine’s arguments. The ACTS project is developing a new generation of technology and standards for automatic differentiation, with particular emphasis on the reverse or adjoint mode of automatic differentiation. The ACTS team is composed of a multidisciplinary group of researchers from MIT, Rice University and Argonne National Laboratory.

The laboratory is exploring how automatic differentiation can be used strategically as a component in more complex numerical algorithms, in particular how the various “cheap derivative” results of automatic differentiation can revolutionize the design of numerical algorithms. We are currently developing algorithms and software for computing directional second order derivatives of functionals with stiff ODEs and DAEs embedded. This approach promises to compute directional second order information at a computational cost proportional to a small multiple of a simulation, independent of the number of parameters involved. We are also developing theory, algorithms and software for adjoint analysis of hybrid dynamic systems, analogous to our previous developments in the forward sensitivity analysis of hybrid dynamic systems.