Efficient neural equalization using multiple nonlinear discrimination functions

Abstract

Conventional neural equalizers treat equalization as an estimation problem. One drawback is that the number of required hidden nodes is large and this results in the high computational complexity and long training period problems. In this paper, we propose a new approach to alleviate this problem. While treating equalization as a classification problem, we divide the received signal space into 2(N) classes (N > 1) for a binary transmitted symbol and apply the discrimination function approach to perform classification. The discrimination functions, which are nonlinear, is realized using a neural network. We also propose a new discriminative learning criterion for the minimum error classification. Simulations show that our approach can greatly reduce the number of hidden nodes leading to shorter training period and lower computational complexity. We can even obtain satisfactory performance without any hidden layers. This yields a very simple equalization structure and is suitable for real-world implementation.