Distributed GA for large system identification problems

TL;DR: A parallel version of a hybrid algorithm of GA and local search is developed for distributed computing, which involves a manager computer running the main algorithm, which distributes data files to many worker computers connected on the network.

Abstract: Non-destructive monitoring of structures may be achieved by system identification to evaluate key parameters. Unfortunately many system identification methods that work for small systems do not necessarily give convergence for large systems. In recent years, the use of genetic algorithms (GA) has shown promising potential for parameter identification of complex systems owing to its many inherent advantages. For large systems involving many degrees of freedom and unknown parameters, the computational effort required by the GA approach may still be prohibitive. The main bulk of computational time lies in the numerous forward analyses that need to be carried out. With rapid advances in computer hardware, especially networking technology, nevertheless, the feasibility of applying the GA approach to large system identification problems has become closer to reality even by using low-cost personal computers. Distributed computing can be easily employed to expedite the GA search, thanks to the high concurrency of the GA approach. In this study, a parallel version of a hybrid algorithm of GA and local search is developed for distributed computing. The implementation involves a manager computer running the main algorithm, which distributes data files to many worker computers connected on the network. Each worker computer carries out the forward analysis with the assigned parameter set and, when completed, sends the output file to the manager computer, Numerical examples are presented to show that this approach is generally workable and robust.