Dynamic structure adaptation in feedforward neural networks-an example of plant monitoring

In the paper artificial neural networks are introduced which are capable of adapting their structure in response to changes in the environment. Feedforward neural networks with multi-layer architecture were trained by modified backpropagation algorithm with forgetting of the connection weights. The applied training algorithm results in a skeleton network structure which can be used for knowledge acquisition. In the authors´ algorithm, the decayed weights are not deleted but fluctuate around zero with a magnitude proportional to the rate of forgetting. Small fluctuations of the weights can grow into a structural evolution in the neural net if properties of the input clusters change. This feature is especially advantageous to on-line system monitoring applications when a rigid neural network structure could lead to mis-interpretation of measurements among dynamically changing conditions. Structural adaptation features and improved generalization capability of the proposed method are illustrated using an example of system state identification in a nuclear reactor