Parallel Particle Swarm Optimization Using Apache Beam

Abstract

The majority of complex research problems can be formulated as optimization problems. Particle Swarm Optimization (PSO) algorithm is very effective in solving optimization problems because of its robustness, simplicity, and global search capabilities. Since the computational cost of these problems is usually high, it has been necessary to develop optimization algorithms with parallelization. With the advent of big-data technology, such problems can be solved by distributed parallel computing. In previous related work, MapReduce (a programming model that implements a distributed parallel approach to processing and producing large datasets on a cluster) has been used to parallelize the PSO algorithm, but frequent file reads and writes make the execution time of MRPSO very long. We propose Apache Beam particle swarm optimization (BPSO), which uses Apache Beam parallel programming model. In the experiment, we compared BPSO and PSO based on MapReduce (MRPSO) on four benchmark functions by changing the number of particles and optimizing the dimensions of the problem. The experimental results show that, as the number of particles increases, MRPSO remains largely constant when the number of particles is small (1000), while BPSO grows slowly and tends to yield better results than MRPSO. As the dimensionality of the optimization problem increases, BPSO can take half the time of MRPSO and obtain better results than it does. MRPSO requires more execution time than BPSO, as the problem complexity varies, but both MRPSO and BPSO are not very sensitive to problem complexity. All program code and input data are uploaded to GitHub.

Keywords

• parallel particle swarm optimization
• Apache Beam
• MapReduce
• swarm intelligence
• big data