Differential Evolution Algorithms For Solving Bilevel Optimization Problems Using Computational Clusters

Bilevel programming problems (BLP5) are a class of optimization problems which are useful to model a series of real-world situations. A BLP is composed of two nested optimization problems and is defined in a hierarchical structure that makes it almost impracticable to solve via classical methods and remarkably expensive to handle via meta-heuristics. Thus, recently we proposed two parallel Differential Evolution implementations to exploit the natural parallelism of metaheuristics and the increasing availability of high-performance computing resources. However, these models were not tested using distributed computing resources such as computational clusters or grids. So, this work provides the first set of experiments for solving BLPs using this type of hardware and clarifies questions about the influence of the communication costs in the performance of different parallel strategies and hardware architectures. Also, speedups were achieved that strongly reinforce the potential of high-performance computing and of the suggested algorithms to enable the solution of increasingly large, complex, practical, and challenging BLPs.