Implicit state observation and control with recurrent neural networks for the bioreactor benchmark problem

We (1993) have recently demonstrated the successful application of dynamic gradient methods to the training of neural network controllers for the bioreactor benchmark, an example of a difficult, nonlinear dynamical process control problem. In this paper, we show that recurrent neural networks can be trained as process controllers for a more difficult version of this benchmark problem in which measurements for only one of the two states are available.