In this paper an efficient kernel affine projection algorithm using dichotomous coordinate descent iterations is proposed. The effectiveness of the proposed algorithm for nonlinear system identification and forward prediction is confirmed by computer simulations.