Asynchronous Parallelization of the Particle Swarm Optimization Algorithm and Its Application to Parameter Estimation in a Large Dimensional Space

Despite its recognized strength in global optimization problems, the high number of objective function evaluations requested by the Particle Swarm Optimization (PSO) method severely limits its application when the computational cost of the objective function calculation is large, as in most actual engineering problems. In order to overcome this deficiency to enable the efficient use of the PSO algorithm in large-scale engineering problems, the parallel computation in clusters appears as an excellent resource. Following this premise, an asynchronous parallelization of the PSO algorithm was developed in this work using the master-slave parallel paradigm and functions of the MPI (Message Passing Interface) library. A set of benchmark tests was conducted to validate and to analyze the performance of the developed method. The developed algorithm, named AIU-PPSO (Asynchronous and Immediate Update Parallel PSO), showed excellent performance, with linear speedup and parallel efficiency higher than 90% for all tested problems. The results were obtained using MIMD parallel computers with distributed memory and 20 Gbits/s Infiniband network. Finally, an actual parameter estimation problem of a population balance model was successfully solved. This problem has a costly objective function and 81 parameters to be estimated.