An Evaluation of Performance Enhancements to Particle Swarm Optimisation on Real-World Data

Swarm Computation is a relatively new optimisation paradigm. The basic premise is to model the collective behaviour of self-organised natural phenomena such as swarms, flocks and shoals, in order to solve optimisation problems. Particle Swarm Optimisation (PSO) is a type of swarm computation inspired by bird flocks or swarms of bees by modelling their collective social influence as they search for optimal solutions. In many real-world applications of PSO, the algorithm is used as a data pre-processor for a neural network or similar post processing system, and is often extensively modified to suit the application. The thesis introduces techniques that allow unmodified PSO to be applied successfully to a range of problems, specifically three extensions to the basic PSO algorithm: solving optimisation problems by training a hyperspatial matrix, using a hierarchy of swarms to coordinate optimisation on several data sets simultaneously, and dynamic neighbourhood selection in swarms. Rather than working directly with candidate solutions to an optimisation problem, the PSO algorithm is adapted to train a matrix of weights, to produce a solution to the problem from the inputs. The search space is abstracted from the problem data. A single PSO swarm optimises a single data set and has difficulties where the data set comprises disjoint parts (such as time series data for different days). To address this problem, we introduce a hierarchy of swarms, where each child swarm optimises one section of the data set whose gbest particle is a member of the swarm above in the hierarchy. The parent swarm(s) coordinate their children and encourage more exploration of the solution space. We show that hierarchical swarms of this type perform better than single swarm PSO optimisers on the disjoint data sets used. PSO relies on interaction between particles within a neighbourhood to find good solutions. In many PSO variants, possible interactions are arbitrary and fixed on initialisation. Our third contribution is a dynamic neighbourhood selection: particles can modify their neighbourhood, based on the success of the candidate neighbour particle. As PSO is intended to reflect the social interaction of agents, this change significantly increases the ability of the swarm to find optimal solutions. Applied to real-world medical and cosmological data, this modification is and shows improvements over standard PSO approaches with fixed neighbourhoods