Abstract Based on analysis of non-linear channel models, a new connectionist model adaptive equalizer is constructed. Comparing with the connectionist model using the Volterra series to extend the input vector space, the number of weights with new structure is reduced significantly. It is shown by simulations that the weight values of the new scheme converge to the optimal values closely for nonminimum phase channels as well as minimum phase channels, if the channel noise is small enough, Testing results of the BER (Bit Error Rate) show that the new adaptive equalizer for non-linear channels is superior to the linear equalizer in the equalization performances.