抄録
Due to the steady progress in technology and the fact that the number of computing resources is increasing, today parallel computing on computer grids or on multi-core systems can significantly reduce the computation time for highly complex modeling, simulation, and optimization problems. Particularly optimizing multi-objective problems is more challenging as there are several objectives which have to be optimized simultaneously. Here, we investigate parallelization of multi-objective algorithms on a set of heterogeneous parallel computers such as a grid. The solution of multi-objective optimization problems is usually a set of solutions represented as an optimal front i.e., none of these solutions can be improved in one objective without getting worse with respect to some other objective. In order to perform the multi-objective optimization on a grid we must parallelize the algorithms. Here, we study two different parallelization paradigms: The Master-Slave and the Island models. In the proposed algorithms for the Island model, every computing resource is responsible for solving one part of the optimal front so that in a collective way all of them obtain the entire optimal solutions. Besides the collective property, we allow the cooperative aspect in the heterogeneous system such that the computing resources indirectly exchange the best found solutions. For the both paradigms, the grid must be represented as a unified resource to the user, who gives the objectives to the system and waits for the optimal solutions. On the other hand, in a heterogeneous system we must make use of all of the resources from very slow to very fast ones. Although the solutions in a multi-objective problem are dependent to each other by the domination relation, the fast processors have to continue working as soon as they are done with their tasks without wasting time in waiting for the slow processors.
