Parallel genetic algorithms on the graphics processing units using island model and simulated annealing

Abstract

To solve a non-deterministic polynomial-hard problem, we can adopt an approximate algorithm for finding the near-optimal solution to reduce the execution time. Although this approach can come up with solutions much faster than brute-force methods, the downside of it is that only approximate solutions are found in most situations. The genetic algorithm is a global search heuristic and optimization method. Initially, genetic algorithms have many shortcomings, such as premature convergence and the tendency to converge toward local optimal solutions; hence, many parallel genetic algorithms are proposed to solve these problems. Currently, there exist many literatures on parallel genetic algorithms. Also, a variety of parallel genetic algorithms have been derived. This study mainly uses the advantages of graphics processing units, which has a large number of cores, and identifies optimized algorithms suitable for computation in single instruction, multiple data architecture of graphics processing units. Furthermore, the parallel simulated annealing method and spheroidizing annealing are also used to enhance performance of the parallel genetic algorithm.