Next: OPTIMA90 Up: Contents Previous: ADOL-F

IMAS: Integrated Modeling and Analysis System

SYNOPSIS:

IMAS derives the tangent-linear and adjoint code to a given Fortran program semi-automatically, using an interpreter approach.

WHOM TO CONTACT:

Andreas Rhodin
GKSS Research Center
Max-Plank Strasse
D-21502 Geesthacht
Germany
Tel.: 49 4152 87 1559
FAX: 49 4152 87 1565
e-mail: rhodin@gkss.de
www: http://w3g.gkss.de/G/imashome

FUNCTIONALITY:

IMAS is an Integrated Modeling and Analysis System for the solution of control problems. It provides algorithms for model inversion (data assimilation, parameter estimation, optimization), sensitivity studies, and the estimation of error covariances. The system derives the tangent-linear and adjoint code to a given Fortran 77 or Fortran 90 program semi-automatically, using an interpreter approach. IMAS is written in Fortran 90 and makes use of object-oriented programming techniques.

HIGHLIGHTS:

Automatic differentiation by an interpreter: The variables of the program must not be declared as REAL variables, but as variables of derived type. The operations (e.g. +,-,*) on this type are overloaded. They perform the additional task of writing a ``tape'' with operations performed and the values of the variables involved in these operations. This tape may be read later by an interpreter that performs the tangent-linear or adjoint operations and thus embodies the tangent-linear or adjoint model, respectively.
Optimization and test algorithms: The original program has to be augmented by statements indicating the control variables and the result variables. These statements have optional parameters in order to pass additional information to the system, for example about constraints, scaling, prior information about the control variables, and data that may be compared to the results in a least square sense. The system provides algorithms that use this informations and perform optimizations, sensitivity studies, and estimation of error covariances.
Modular design: IMAS has a modular design in order to allow the easy exchange of the forward model and of the basic data types. The respective modules can be exchanged independently from the other parts of the code. Thus the derived data types and the basic operations can easily be adapted to a specific model. The nonlinear model can be changed without recompiling the other modules, which are provided in a library. Advanced data types for large applications or special purposes (CFD-codes) are available or under development.

Next: OPTIMA90 Up: Contents Previous: ADOL-F