Particle Swarm Optimization And Its Application In Pattern Recognition

This paper first discusses the original algorithm, the history of PSO and analyzestheconvergenceandits conditions ofthesimplified versionofPSO.Then,it applies thePSO to multi-peak searching problem and trains the neural network with a new class ofPSO.Themaincontributionsofthispaperareasfollows:(1)A class ofdynamic-populationparticleswarm optimizationforsearchingpeaksof some multi-peak functions is proposed. This algorithm transforms all peaks ofmulti-peak problems into those peaks equally high by functional transformation, inorder to find all peaks with the same probability. During the searching the size ofparticle swarm could be adjusted to get any initial size of swarm. So it could solve theproblem of determining swarm size because the number of peaks of the givenmulti-peak function could not be obtained in standard particle swarm optimization. Theexperiments manifest that the algorithm can search peaks of function as much aspossible.(2) A class of excellent individual sub-population based on particle swarmoptimization for searching peaks of some multi-peak functions is proposed. Thisalgorithm transforms all peaks of multi-peak problems into those peaks equallyhigh byfunctional transformation, in order to find all peaks with the same probability. Duringthe searching, to delete some particles in particle swarm of every generation, whichbeside the peaks that have been found, such can avoid searching for the same peak anddecrease the generation number of particle swarm for searching. The experimentsmanifestthatthealgorithmcansearchpeaksoffunctionasmuchaspossible.(3) In accordance with problems of the too-strong stochastic characteristic andslow convergence speed of the traditional particle swarm optimization. Two newposition update equations are proposed by using the strategy of extrapolation inMathematics. Thus, a new class of induction-enhanced particle swarmoptimization(IEPSO)isgiven.ThenewalgorithmisbetterthanthetraditionalPSOwithboth a better stability and a steady convergence, especially for some function withmulti-dimension. Then the IEPSO is proposed to train neural networks. The experimentmanifest that the IEPSO can help the networks overcome the weakness which may beproducedbytheback-propagationalgorithm.