Implementing and evaluating parallel evolutionary algorithms in modern GPU computing libraries

Abstract

In this paper, we describe FFEAT - a library for GPU-based implementation of evolutionary algorithms based on Torch. We discuss limitations of GPU computing and how they affect implementations of evolutionary algorithms and other population-based heuristics.

Using FFEAT, we implement a number of different types of nature inspired algorithms, including evolutionary algorithms, evolution strategies, and particle swarm optimization. We investigate the performance of such algorithms in a number of benchmarks and with varying algorithm settings.

We show that in some cases, we can obtain an order of magnitude speed-up by running the algorithm on a GPU compared to running it on a CPU.

Keywords

• evolutionary algorithms
• GPU computing
• parallelization