The lack of human interaction in automatic reflection tomography leads to a larger percentage of ``bad'' data points. The number of data points associated with events with spurious moveouts (such as multiples and converted waves) can be minimized by intelligently controlling the semblance scanning range. The effect of the bad data points can be limited by replacing the standard L2 norm solution with a norm closer to L1 by reweighted least-squares. By replacing the standard constant parameter with a diagonal operator, areas with large errors in moveout can be highly regularized with minimal effect on areas with more reliable moveout information. This methodology is applied to a complex 2-D dataset.