Robust Modeling With Erratic Data

Abstract An attractive alternative to least-squares data modeling techniques is the use of absolute value error criteria. Unlike the least-squares techniques the inclusion of some infinite blunders along with the data will hardly affect the solution to an otherwise well-posed problem. An example of this great stability is seen when an average is, determined by using the median rather than the arithmetic mean. Algorithms for absolute error minimization are often approximately as costly as least-squares algorithms; however, unlike least-squares, they naturally lend themselves to inequality or bounding constraints on models.