A digital communication channel equalizer using a Kalman-trained neural network

This paper presents a neural network-based equalizer for a digital communication system. In this equalizer, the neural network is adapted to the optimum decision boundaries as defined by the channel and noise characteristics. This problem has already been considered by Gibson et al.(1991), but for more complex decision boundaries, the simple LMS backpropagation training rule used in their paper leads to excessively large numbers of training steps. In this paper, the neural network is trained using the extended Kalman algorithm, which has been described by Iiguni, Sakai, and Tokumaru (1992). Here a different sigmoid nonlinearity is used which is easier to compute. It is shown that a significant reduction in the number of training steps is obtained. Finally, a simplified and generalized derivation of the Kalman training algorithm is presented