A hybrid structure for adaptive fixed weight recurrent networks

Due to the evolution of the underlying physical process, a correct model can transform into an erroneous one. We therefore propose a method which overcomes this problem by adapting the network along the way. Our method (clustered error injection) is based (a) on the ability of real-time recurrent learning networks to form clustered network structures and (b) on the error injection principle. The actual model error is fed back into the network as an input. This improves the model performance by adapting it to a changing environment. This technique is tested on two examples, a mathematical modelling problem and a real-life problem from the chemical process industry