Order statistic filters form a subclass of the stack filters. The design of stack filters is computationally prohibitive for relatively large windows. A possible approach to overcome this difficulty is to constrain design to a smaller class. We propose an algorithm for the design of optimal order statistic filters from training data, and show how it can also be applied to the design of optimal weighted order statistic filters. Some examples that illustrate application of the proposed algorithm are presented, with considerations on training time and precision of the designed filter.