Efficiency of particle swarm based method parallelization in optimizing training of neural networks

Larionov Vyacheslav Sergeevich, Maleev Oleg Gennadievich

Subject of research: The possibility of using parallel computations of the particle swarm optimization algorithm (PSO) in classification problem using feedforward neural network. Three datasets with various lengths, different number of instances, input features and weights numbers were used for calculations; neural network training was carried out for 500 epochs in 50 independent program runs using different processors number for parallelization. Cross-entropy loss function was chosen for the penalty function in order to estimate performance of the proposed methods. The advantage of using parallel computations with the PSO was shown on the designed networks analysis results.

Keywords:neural network, optimization, PSO, PAPSO, Broadcast PPSO.