Development Page--Not for Official Use

Learning or Training Phase

A distinctive feature of an ANN is the models ability to learn. An ANN learns by adaptively updating the synaptic weights that characterize the strength of the connections between neurons. The weights are updated according to the information extracted from each successive training pattern. Typically, the optimal weights are obtained by minimizing or maximizing certain "energy" functions. For example, a common criterion is to minimize the least-squares-error between the teacher value and the actual output value.

The DNR ANN model was trained using the Levenberg-Marquardt algorithm as implemented within the Matlab Neural Network Toolbox. This method of nonlinear estimation demonstrates practical reliability in that it is capable of converging promptly from a wider range of initial guesses than other typical methods.

An interesting research project is in development testing EngineeredTrainingSets or the variability due to artificially created data sets based on severe weather patterns. In theory, these training sets will create a difficult learning environment that should allow the Artificial Neural Networks to adapt to severe weather based on their new ability to associate rapid water level changes due to wind forcing of the worst local weather.

Return to "What Exactly is an Artificial Neural Network?"

Return to "Neural Network Application to Water Level Forecasting"