Generating adaptive models of dynamic systems with recurrent neural networks
T. Catfolis · 2002
Presents a method for building adaptive neural network models based on the real-time recurrent learning (RTRL) algorithm developed by Williams and Zipser (1989). The author introduces the error injection method as adaptation principle. This method consists of feeding back the model error as input to the network what causes the model to react to it. The main advantages of this method are a higher stability, and a better and faster model compared to networks using only the RTRL algorithm as adaptation rule. The author demonstrates the properties of this technique with a mathematical example and with an example based on a bioreactor model.>