engleski

Analysis of Iterative Learning Algorithms for the Multilayer Perceptron Neural Network

In this paper, a comparison of different algorithms used in the training of a multilayer feedforward neural network is presented. Tested algorithms, which are of the first and the second order, include both local and global adaptation techniques. Prediction of nonlinear dynamic Glass-Mackey system is used as a benchmark problem. To improve training speed and efficiency, bipolar sigmoidal activation function with adaptive gain parameter is used. Furthermore, modification of random weight initialization is proposed.

static neural network; adaptive activation function; prediction; nonlinear chaotic system

nije evidentirano