Abstract
In this article we demonstrate the identification of a nonlinear,
dynamic process with recurrent neural structures. The employed
network-structure is a Recurrent Multilayer Perceptron (RMLP), which
combines feedforward-- and recurrent architectures. We will show that
RMLPs are capable of learning the temporal behavior and characteristic
of an arbitrary, nonlinear, dynamic process. Apart from conventional
gradient-based algorithms, a sophisticated statistical method has been
considered for this challenging task - Global Extended Kalman Filtering
(GEKF). This powerful algorithm yields neural structures with a
significantly better performance, compared to conventional
gradient-based approaches. The new element in this work is the
application of the GEKF-Algorithm for recurrent neural structures,
which are employed in the identification of nonlinear, dynamic
processes. In order to supervise the quality of network-training,
appropriate performance-indexes for neural identification are
introduced. The distribution of the Moving Average Squared Error (MASE)
is employed as an objective optimality-criterion, in order to survey
the actual performance of recurrent neural structures during training.
Users
Please
log in to take part in the discussion (add own reviews or comments).