A high performance memetic algorithm for extremely high-dimensional problems

GPU parallelization strategies for metaheuristics: a survey

“…In the same context, population based metaheuristics have been widely exploited to solve high dimensional continuous problems. Besides iteration level, in most of works, as [16,36,42,49,[51][52][53], dimensions of individuals are computed in parallel (solution level), which increases the acceleration of the algorithms. Quality has been preserved, and even improved as it is the case in [36,42,44,49,90].…”

“…Finally, the common point of the works in designing a GPU based algorithm is trying to maximize data parallelism. Works like [16,21,42,50,51] represent solutions as thread blocks where threads evaluate dimensions, or perform specific operators, such as crossover, mutation or to generate the neighborhood. Moreover, GPU impact becomes clear when big instances are addressed.…”

“…Other evolutionary algorithms: A new design of a memetic algorithm, called MA-SW-Chains is presented in [51] for GPU architecture. Its main idea is to combine a steady-state genetic algorithm (SSGA) [86] with a local search procedure, called Solis Wets search method [70].…”

“…Its main idea is to combine a steady-state genetic algorithm (SSGA) [86] with a local search procedure, called Solis Wets search method [70]. According to [51], the GPU accelerated MA-SW chains has two natural sources of parallelism: the number of individuals in the population (iteration level), and the number of variables of each individual (solution level). The steps that have been parallelized in the algorithm are: evaluation of the fitness function, adaptation of the crossover operator to the GPU, optimization of the local search, random number generation process, and population sorting.…”

Cascaded evolutionary algorithm for nonlinear system identification based on correlation functions and radial basis functions neural networks

“…Moreover, by the application of parallel programming to software development such as GPUs or computer clusters, we may dramatically alleviate the total time to build a model, as the algorithm structure is easily parallelized. This has been already done with the GA [61] and other EAs [62]. On the other hand, the final model is a NN and thus a massively parallel structure, that makes it very convenient to be employed in embedded hardware [63] and provide an estimate almost instantaneously.…”