Particle Swarm Optimization Based Determination of Learning Parameters in Artificial Neural Networks with Backpropagation Learning Algorithm
Emre Çelik Nihat Ozturk Adem Dalcalı
AbstractIn this paper, a particle swarm optimization (PSO) algorithm is introduced to determine learning parameters required for the backpropagation (BP) learning algorithm, which is used for training of a feed-forward neural network (FFNN). PSO algorithm utilized within the paper works slightly different compared to conventional PSO (CPSO) algorithm in such a way that each particle adjusts its position based on the best midposition of all particles and its group’s previous best. The major reason of such a change is to enhance the performance of CPSO algorithm, which is explained in detail in the study suggested by Tamer, S and et.al. To test the proposed method, a FFNN with three layers is designed for function interpolation. Learning parameters of the designed NN are determined by both conventional error and trial method and the proposed method. Afterwards, using these two groups of learning parameters, the NN is trained and tested under the same conditions. According to the test results, learning parameters determined by the PSO provide a better performance and interpolating capability for the NN than those determined by the conventional method.