A method and apparatus are disclosed for improving channel equalization and level learning in a data communication system. The disclosed equalizer training process separately updates the feed forward filter (FFF) and the level adapter, to gain additional improvements in the training of the feed forward filter (FFF). The multi-step equalizer training process initially trains the feed forward filter (FFF) using a two-level signal EQTR(n) having an ideal value (Step One) to help converge the feed forward filter (FFF) to a certain level. Once the feed forward filter (FFF) reaches a certain level of convergence, the training circuitry is reconfigured during step two of the equalizer training process, to evaluate and update the actual level of the signal EQTR(n), to compensate for the channel. The determined weighting factors are applied to a low pass filter and the actual level of the signal EQTR(n), B(n), is calculated. Once the actual level of the signal EQTR(n), B(n), has been calculated, the actual level of the signal EQTR(n), B(n), is applied to the level adapter, and the level adapter is no longer updated by reconfiguring the training circuitry to remove the error signal, err(n), inputs to the level adapter. During step three, the feed forward filter (FFF) continues to be updated and fine-tuned by the error signal err(n). Since the level of the signal EQTR(n) is the actual value, B(n), the performance of the feed forward filter (FFF) is improved. Once the equalizer training process is complete, the feed forward filter (FFF) is fixed. The improved training of the feed forward filter (FFF) allows the structure of level learning process to be simplified, with the training circuitry removed and the feed forward filter (FFF) fixed, where each level will be divided into six phases and processed individually.