Konferensartikel

MATLAB® Neural Network (NN) Toolbox can handle both static and dynamic neural networks. Using MATLAB® NN Toolbox with recurrent neural networks is not straight forward. We present a Dynamic Arti?cial Neural Network (DANN) MATLAB toolbox capable of handling fully connected neural networks for time-series analysis and predictions. Three different learning algorithms are incorporated in the MATLAB DANN toolbox: Back Propagation Through Time (BPTT) an of?ine learning algorithm and two online learning algorithms; Real Time Recurrent Learning (RTRL) and Extended Kalman Filter (EKF). In contrast to existing MATLAB® NN Toolbox, the presented MATLAB DANN toolbox has a possibility to perform the optimal tuning of network parameters using grid search method. Three different cases are used for testing three different learning algorithms. The simulation studies con?rm that the developed MATLAB DANN toolbox can be easily used in time-series prediction applications successfully. Some of the essential features of the learning algorithms are seen in the graphical user interfaces discussed in the paper. In addition, installation guide for the MATLAB DANN toolbox is also given.

dynamic arti?cial neural network (DANN), back propagation through time (BPTT), real-time recurrent learning (RTRL), extended Kalman filter (EKF), time series