RECURRENT NEURAL NETWORK TRAINING BY AN OPTIMAL CONTROL ALGORITHM

摘要

Motivated by the methodologies presented in Farotimi et al., we consider a recurrent neural network as a plant to be controlled according to a prescribed performance index using optimal control techniques. The parametric weights of the system are considered as the control variables. The objective of our paper is to derive algorithms for the training of parameters of the dynamical system such that the system output approximates a given one-dimensional time series under the condition that the series is pre-processed by a suitable data transformation. The sunspots and the Mackey-Glass series are used as test sequences. Algorithms are then derived for computer experiments with Matlab. We use numerical iterative techniques to optimize the system parameters such that the least-squared prediction errors are minimized. One of the novelties in this paper is to derive algorithms applicable to time series simulation based on a recurrent neural network model and using optimal control techniques.