To improve the performance of nonlinear RF power amplifiers, either in terms of spectrum control or constellation quality, digital predistortion has become an attractive alternative to the simple approach of `backing off' the amplifier into a quasi-linear region of operation. This allows more efficient use of the amplifier in both RF output and DC power consumption. In this paper, we formulate a sequentially-adaptive algorithm for training a complex baseband predistorter that minimizes mean-square-error between the complex input signal and the eventual amplifier output. The algorithm fits within the direct learning category of adaptive filters, and extends other adaptive filter formulations to the complex signal and nonlinear system domain. We study performance of the algorithm on several test models for amplifiers, in terms of normalized mean square error, output power spectrum and constellation quality.