Particle swarms with dynamic topologies and conservation of function evaluations

Abstract

This paper proposes a general framework for structuring dynamic Particle Swarm populations and uses a conservation of function evaluations strategy to increase the convergence speed. The population structure is constructed by placing the particles on a 2-dimensional grid of nodes, where they interact and move according to simple rules. During the running time of the algorithm, the von Neumann neighborhood is used to decide which particles influence each other when updating their velocity and position. Each particle is updated in each time-step but they are evaluated only if there are other particles in their neighborhood. A set of experiments demonstrates that the dynamics imposed by the structure provides a more consistent and stable behavior throughout the test set when compared to standard topologies, while the conservation of evaluations significantly reduces the convergence speed of the algorithm. Furthermore, the working mechanisms of the proposed structure are very simple and, except for the size of the grid, they do not require parameters and tuning.