Particle swarm optimisation for object classification
H. Evans, Miaoru Zhang · 2008
This paper describes a new approach to the use of particle swarm optimisation (PSO) for object classification problems. Instead of using PSO to evolve only a set of good parameter values for another machine learning method for object classification, the new approach developed in this paper can be used as a stand alone method for classification. Two new methods are developed in the new approach. The first new PSO method treats all different features equally important and finds an optimal partition matrix to separate a data set into distinct class groups. The second new PSO method considers the relative importance of each feature with the noise factor, and evolves a weight matrix to mitigate the effects of noisy partitions and feature dimensions. The two methods are examined and compared with a popular method using PSO combined with the nearest centroid and another evolutionary computing method, genetic programming, on three image data sets of increasing difficulty. The results suggest that the new weighted PSO method outperforms these existing methods on these object classification problems.