Distributed solution concepts in simulation-based decision support

Abstract

Decision support mostly implies the solution of domain-specific optimization problems. Whether in management, economics or in engineering sciences the simulation of complex systems (product design as well as production planning or supply chain management) plays an important role. In many cases, these problems involve numerically complex and therefore computationally expensive simulations. These simulation-based constrained nonlinear optimization and control problems can be characterized as nonconvex and non-smooth regarding objective and constraint functions. This causes several specific tasks for the numerical solution: (1) determining the global solution, (2) handling the non-availability of analytical derivatives, and (3) integration of black-box type simulation-packages with the optimization solver. The solution of this kind of optimization problems is based on coupling optimization with simulation; that means managing the communication and synchronization between the simulation and optimization systems. Due to the large computational effort in solving simulation-based optimization problems, the computation time has to be reduced significantly. This can be done by using parallel and distributed concepts in the solution process. For the nonsequential solution of simulation-based optimization problems on a network of workstations an adequate software environment is developed. A distributed optimization algorithm is introduced and the parallel performance is discussed. A corresponding software architecture is described and results from applying the distributed system to optimization and control problems in groundwater management are presented.