Abstract:The paper investigates the application of a simple nonlinear structure to the problem of adaptive channel equalisation. Based on the Bayes decision rule, it is shown that the optimal equalisation solution is an inherently nonlinear problem and, therefore, it is desired to incorporate some degree of nonlinearity in the design of equaliser structure. The approximate realisation of the optimal equalisation solution is implemented using a polynomial-perceptron architecture and simulation results are included to support the theoretical analysis.
IntroductionCommunications channel equalisation is concerned with the reconstruction of digital signals that have been passed through a dispersive channel and then corrupted with additive noise. Traditional techniques for solving this equalisation problem are based on linear finite filters. Adaptive linear equalisers are robust and can easily be implemented. The operation of an equaliser at each sample instant is typically based on a finite number of channel observations and decisions are usually made on a symbol-by-symbol base. Even under this classical information constraint, it has been shown that channel equalisation is an inherently nonlinear problem [4] regardless of whether a channel is minimum or nonminimum phase. Nonlinear structures are therefore required to achieve fully or near optimal performance. Gibson et al.[4] proposed a nonlinear equaliser structure based on the multilayer perceptron and demonstrated its superior performance over the linear equaliser. The multilayer perceptron has a very general ability of nonlinear decision making and, theoretically, a multilayer perceptron equaliser with sufficient size can realise the optimal performance. There are, however, some practical difficulties associated with this highly nonlinear structure that require further investigation. The selection of architecture and parameter values for the multilayer perceptron equaliser is mainly by experiment. The training [7], and training times are typically very long. Although the use of recursive Gauss-Newton algorithms [2, 91 can significantly improve the convergence properties of the multilayer perceptron equaliser, these algorithms require more computation at each recursion and will have difficulties in meeting the real-time requirements of high-speed data transmission where adaptive equalisation is mostly needed.In this paper an alternative nonlinear equaliser structure is examined. Using the Bayes decision rule, it is shown that the optimal equalisation solution is highly nonlinear, a result identical to that derived in [4] by a different approach. An old technique, namely polynomial approximation, is then employed as a means of approximately realising the optimal solution. This leads to a polynomial-perceptron equaliser that is theoretically more tractable compared with the multilayer perceptron equaliser as the filter parameters are almost linear with respect to the output error. Simple simulation examples are included to compare the performance of this polynomia...