A comparison of nonlinear Kalman filtering applied to feed-forward neural networks as learning algorithms

Abstract

In this article we present an application of Kalman filtering in Artificial Intelligence, where nonlinear Kalman filters were used as a learning algorithms for feed-forward neural networks. In the first part of this article we have examined two modern versions of nonlinear filtering algorithms i.e. Unscented Kalman Filter and Square Root Central Difference Kalman Filter. Later, we present performed experiments, where we have compared UKF and SRCDKF with an reference algorithm i.e. Error Backpropagation being the most popular neural network learning algorithm. To prove filters high learning abilities in case of noisy problems, we have used a noisy financial dataset during the experiments. This dataset was selected due to uneasily separable classes subspaces. The results of experiments, presented in the last part of this paper, show greater accuracy for nonlinear Kalman filters that overperformed popular Error Backpropagation learning algorithm.