Parallel algorithms for pso

Similar to most stochastic optimization algorithms, the PSO algorithm can be computationally intensive when the computation of the adaptation value evaluation function is relatively large. In order to solve the problem, researchers proposed the parallel PSO algorithm. Similar to the parallel genetic algorithm, the parallel PSO algorithm can have three parallel population models: master-slave parallel model, island population model, and adjacency model.

Schutte uses a synchronized implementation, where the adaptation values of all points in a generation are computed before moving on to the next generation. This parallel approach, although simple to implement, often results in poor parallel efficiency. Therefore, an asynchronous approach to parallelization has been proposed, which can improve the parallel performance of the PSO algorithm with little impact on the numerical accuracy. Both approaches use the master-slave parallel model, in which the asynchronous approach is more coupled in the solution and more likely to generate communication bottlenecks.

Baskar proposed a concurrent PSO algorithm with two subpopulations evolving in parallel, where one subpopulation adopts the original PSO algorithm, and the other subpopulation adopts the PSO algorithm based on the fitness-value-distance-ratio (FDR-PSO); information is exchanged frequently between the two subpopulations. And El-Abd studied the collaborative PSO algorithm using local neighborhood version in subpopulations, and investigated the multiple ways of information exchange and their effects on the performance of the algorithm. Huang-Fang proposed a parallel PSO algorithm based on the island population model and introduced a centralized migration strategy to improve the solution efficiency as well as the early convergence phenomenon.

Li proposes a parallel algorithm for delayed exchange of information belonging to the neighbor-joining model, which improves the speed but may make the quality of the solution worse.