Speeding up DSMGA-II on CUDA platform
Sung-Chi Li, Tian–Li Yu · Proceedings of the Genetic and Evolutionary Computation Conference · 2017
This paper proposes two CUDA based implementations to speed up the model building process for DSMGA-II, which has shown superior optimization ability to hBOA and LT-GOMEA on various benchmark problems. The first implementation is lossless, which is algorithmically identical to the original version. The second implementation is lossy, which sacrifices some accuracy for further speedup. On several commonly used benchmark problems, the proposed implementations are stable. As the problems become larger, the amount of speedup increases accordingly. The lossless scheme speeds up the first part of model building for more than ten times; the lossy implementation further speeds up the second part of model building for more than 400 times on a 600-bit folded-trap problem. The limitation of such implementations are also discussed in detail in this paper.