Cellular genetic algorithm on graphic processing units

Abstract

The availability of low cost powerful parallel graphic cards has estimulated a trend to implement diverse algorithms on Graphic Processing Units (GPUs). In this paper we describe the design of a parallel Cellular Genetic Algorithm (cGA) on a GPU and then evaluate its performance. Beyond the existing works on masterslave for fitness evaluation, we here implement a cGA exploiting data and instructions parallelism at the population level. Using the CUDA language on a GTX-285 GPU hardware, we show how a cGA can profit from it to create an algorithm of improved physical efficiency and numerical efficacy with respect to a CPU implementation. Our approach stores individuals and their fitness values in the globalmemory of the GPU. Both, fitness evaluation and genetic operators are implemented entirely on GPU (i.e. no CPU is used). The presented approach allows us benefit from the numerical advantages of cGAs and the efficiency of a low-cost but powerful platform. © 2010 Springer-Verlag Berlin Heidelberg.