Nonlinear mappings based on particle swarm optimization

Nonlinear mapping methods that minimize the Sammon stress based on particle swarm optimization (PSO) are proposed. The task considered is the mapping of the codebook vectors generated by the neural gas (NG) network onto a two-dimensional space. Three methods are explored: the direct application of the traditional PSO, the initialization of PSO with TOPNG, and a dynamically growing PSO. These methods are compared with the Sammon´s mapping and TOPNG in terms of the Sammon stress and the topology preservation measure q_m. The best results are obtained when PSO is initialized with TOPNG.