Updating Random Weights in Artificial Neural Networks based on Particle Swarm Optimization

All weights are usually updated in every iteration of artificial neural networks training phase and this causes to lower convergence speed. Salvetti and wilamowski introduced random weight updating process in feed-forward artificial neural networks to improve probability and convergence speed. In this paper, we tested updating of random weight on an artificial neural network based on particle swarm optimization. Unlike updating of weights in artificial neural networks which all weights are updated, only some of weights are randomly updated by particle swarm optimization in every iteration. Results of testing the proposed method on data of circle-in-square problem show that replacing regular updating of all weights with random updating of weights averagely decreases accuracy by about 0.3% in addition that it improves convergence speed. The previous similar method has decreased accuracy of 2.3% in the best case.

کلیدواژه ها:

ANN ، PSO ، Updating random weights