• multimodal function optimization.
• niching
• Particle Swarm Optimization
• swarm intelligence

The Particle Swarm Optimization (PSO) algorithm, like many optimization algorithms, is designed to find a single optimal solution. When dealing with multimodal functions, it needs some modifications to be able to locate multiple optima. In a parallel with Evolutionary Computation algorithms, these modifications can be grouped in the framework of Niching.
In this thesis, we present a new approach to niching in PSO that is based on clustering particles to identify niches. The neighborhood structure, on which particles rely for communication, is exploited together with the niche information to perform parallel searches to locate multiple optima. The clustering approach was implemented in the k-means based PSO (kPSO), which employs the standard k-means clustering algorithm. We follow the development of kPSO, starting from a first, simple implementation, and then introducing several improvements, such as a mechanism to adaptively identify the number of clusters.
The final kPSO algorithm proves to be a competitive solution when compared with other existing algorithms, since it shows better performance on most multimodal functions in a commonly used benchmark set.