Existing alignment algorithms all assume that the alignment signal is symmetrical about the correct center position. When the signal becomes asymmetrical, these algorithms inevitably result in alignment error. We describe a general approach to align accurately on asymmetrical signals. This is achieved by incorporating learning and utilization of a priori information. The proposed algorithm looks at some sample alignment signals with known centers. The latter are provided by metrology data or some other means. The algorithm builds a linear space model of the asymmetry that is present in the sample signals. It then uses the built model to extract the symmetrical part of alignment signals that come from the same, well-controlled process. The extracted nearly symmetrical signal is then used to determine the alignment position. A detailed algorithm is provided for each of the three steps. Computer simulation implementing the algorithms shows that the alignment performance, both in terms of the mean and variance of the alignment error, is significantly improved compared to two examples of alignment algorithm that do not incorporate learning. The two example alignment algorithms used for comparison are phase detection and center of mass detection. A physical interpretation of the linear asymmetry model is provided.