Simplifying Particle Swarm Optimization

TLDR: It is found that not only the PSO method and its simplified variant have comparable performance for optimizing a number of Artificial Neural Network problems, but also the simplified variant appears to offer a small improvement in some cases.

Abstract: The general purpose optimization method known as Particle Swarm Optimization (PSO) has received much attention in past years, with many attempts to find the variant that performs best on a wide variety of optimization problems. The focus of past research has been with making the PSO method more complex, as this is frequently believed to increase its adaptability to other optimization problems. This study takes the opposite approach and simplifies the PSO method. To compare the efficacy of the original PSO and the simplified variant here, an easy technique is presented for efficiently tuning their behavioural parameters. The technique works by employing an overlaid meta-optimizer, which is capable of simultaneously tuning parameters with regard to multiple optimization problems, whereas previous approaches to meta-optimization have tuned behavioural parameters to work well on just a single optimization problem. It is then found that not only the PSO method and its simplified variant have comparable performance for optimizing a number of Artificial Neural Network problems, but also the simplified variant appears to offer a small improvement in some cases.