A study of gradient based particle swarm optimisers

Abstract

Gradient-based optimisers are a natural way to solve optimisation problems, and have long been used for their efficacy in exploiting the search space. Particle swarm optimisers (PSOs), when using reasonable algorithm parameters, are considered to have good exploration characteristics. This thesis proposes a specific way of constructing hybrid gradient PSOs. Heterogeneous, hybrid gradient PSOs are constructed by allowing the gradient algorithm to optimise local best particles, while the PSO algorithm governs the behaviour of the rest of the swarm. This approach allows the distinct algorithms to concentrate on performing the separate tasks of exploration and exploitation. Two new PSOs, the Gradient Descent PSO, which combines the Gradient Descent and PSO algorithms, and the LeapFrog PSO, which combines the LeapFrog and PSO algorithms, are introduced. The GDPSO represents arguably the simplest hybrid gradient PSO possible, while the LeapFrog PSO incorporates the more sophisticated LFOP1(b) algorithm, exhibiting a heuristic algorithm design and dynamic time step adjustment mechanism. The strong tendency of these hybrids to prematurely converge is examined, and it is shown that by modifying algorithm parameters and delaying the introduction of gradient information, it is possible to retain strong exploration capabilities of the original PSO algorithm while also benefiting from the exploitation of the gradient algorithms.