OpenCL Implementation of PSO Algorithm for the Quadratic Assignment Problem

Abstract

Abstract. This paper presents a Particle Swarm Optimization (PSO) algorithm for the Quadratic Assignment Problem (QAP) implemented on OpenCL platform. Motivations to our work were twofold: firstly we wanted to develop a dedicated algorithm to solve the QAP showing both time and optimization performance, secondly we planned to check, if the capabilities offered by popular GPUs can be exploited to accelerate hard optimization tasks requiring high computational power. We were specifically targeting low-cost popular devices, with limited capabilities. The paper discusses the algorithm and its parallel implementation, as well as reports results of tests.

Multi-swarm PSO Algorithm for the Quadratic Assignment Problem: A Massively Parallel Implementation on the OpenCL Platform

“…Their results are gathered in Table 1 and Table 2. The selection of algorithm configuration parameters (c 1 , c 2 and c 3 factors, as well as the kernels used) was based on previous results published in [33]. In all cases the second target S x aggregation kernel was applied (see Algorithm 1), which in previous experiments occurred the most successful.…”

“…The gap is between 0% and 6.4% for the biggest case tai150b. We have repeated tests for tai60b problem to compare the implemented multi-swarm algorithm with the previous single-swarm version published in [33]. Gap values for the best results obtained with the single swarm algorithm were around 7%-8%.…”

“…In our recent work [33] we have developed PSO algorithm for the Quadratic Assignment Problem on OpenCL platform. The algorithm was capable of processing one swarm, in which particles shared information about the global best solution to update their search directions.…”

“…Such organization of computations involves inevitable overhead related to data transfer between the host and the GPU device. The time efficiency test reported in [33] showed clearly that the benefits of a parallel execution platform can be fully exploited, if processed populations are large, e.g. if they comprise several hundreds or thousands particles.…”

Parallel Hierarchical Genetic Algorithm for Scattered Data Fitting through B-Splines

“…There are techniques and methods that are intrinsically parallel, and can be implemented by taking advantage of new technologies for data processing, such as graphics processing units (GPUs) and multiprocessor computers, as well as tools developed for this purpose. The parallelization of methods and techniques has been widely developed in different knowledge fields showing significant improvements in response times for computationally expensive problems (see [18][19][20]). This work is focused on hierarchical genetic algorithms and its implementation and application as a parallel architecture for curve fitting.…”