Adaptive local Radial Basis Function network

nonstationary channel equalisation problem Eng-Siong Chng y 1, Howard Yang z, Herbert Wiklickyz y Institute of Systems Science, National University of Singapore, Singapore 11957. Email : [email protected] z Brain Information Processing Group, Frontier Research Programme, RIKEN. 2-1 Hirosawa, Wako-shi, Saitama 351-01, JAPAN. Abstract| The computational requirement to implement the optimal Bayesian symboldecision equaliser using RBF network [1] can be very high as the full RBF Bayesian solution usually requires a large number of centres. To reduce the implementation complexity, we propose to use a subset number of the full RBF network's centres to generate a subset equaliser. The centres to be selected for the subset equaliser are those that have their Euclidean distance close to the equaliser's current input vector. Our results show that the number of centres can be greatly reduced without signi cant degradation in classi cation performance.