The particle swarm optimization (PSO) algorithm is a stochastic, population-based optimization technique influenced by social dynamics. It has been shown that the performance of the PSO algorithm can be greatly improved if the control parameters are appropriately tuned. However, the tuning of control parameter values has traditionally been a time-consuming, empirical process followed by statistical analysis. Furthermore, ideal values for the control parameters may be time-dependent; parameter values that lead to good performance in an exploratory phase may not be ideal for an exploitative phase. Self-adaptive algorithms eliminate the need to tune parameters in advance, while also providing real-time behaviour adaptation based on the current problem.
This thesis first provides an in-depth review of existing self-adaptive particle swarm optimization (SAPSO) techniques. Their ability to attain order-2 stability is examined and it is shown that a majority of the existing SAPSO algorithms are guaranteed to exhibit either premature convergence or rapid divergence. A further investigation focusing on inertia weight control strategies demonstrates that none of the examined techniques outperform a static value. This thesis then investigates the performance of a wide variety of PSO parameter configurations, thereby discovering regions in parameter space that lead to good performance. This investigation provides strong empirical evidence that the best values to employ for the PSO control parameters change over time. Finally, this thesis proposes novel PSO variants inspired by results of the aforementioned studies.