Parallelizing evolutionary algorithms for clustering data

TL;DR: In this paper, a novel approach, based on data decomposition, for parallel computing of the fitness function is proposed, where both the learning set and the population of the evolutionary algorithm are distributed among processors.

Abstract: In the paper the problem of using an evolutionary algorithm to partition a dataset into a known number of clusters is considered. A novel approach, based on data decomposition, for parallel computing of the fitness function is proposed. Both the learning set and the population of the evolutionary algorithm are distributed among processors. Processors form a pipeline using the ring topology. In a single step each processor computes the local fitness of its current subpopulation while sending the previous subpopulation to the successor and receiving next subpopulation from the predecessor. Thus it is possible to overlap communication and computation using non-blocking MPI routines. Our approach to parallel fitness computation was applied to differential evolution algorithm. The results of initial experiments show, that for large datasets the algorithm is capable of achieving very good scalability.