Group-based stochastic scaling for PSO velocities
E.T. van Zyl, Andries Petrus Engelbrecht · 2016
This paper examines a group-based approach to stochastically scaling the cognitive and social components of a particle swarm's velocities. Usually, such scaling is done by generating a random vector and then multiplying in a component-wise fashion. Instead, the problem's decision variables are divided into a number of groups and every group is scaled with a random number. Three different grouping strategies are provided: fixed group number, decreasing group number and increasing group number. These grouping strategies are compared with a standard PSO and amongst one another on a well known suite of high-dimensional benchmark functions. The proposed grouping strategies were all more effective or equal to scaling the velocity components in the standard way. It was found that linearly increasing the number of groups that the decision variables are divided into significantly outperforms the standard update method and all of the other grouping strategies considered. A detailed discussion of the results obtained and a brief empirical investigation of any parameters introduced by the grouping strategies are also provided.