Robust Backpropagation and Statistical Process Control

Abstract: It is well known that outliers have a major deleterious effect on statistical methods based on ordinary least squares. Robust techniques have been developed in order to combat these effects. Since neural networks are related to regression algorithms, outliers can produce the same the same types of problems in this case as well. Chen and Jain (1994) developed a robust backpropagation procedure to downweigh the effect of outliers in a manner similar to that used in present day robust statistics. The purpose of this paper is to test this algorithm for use in Statistical Process Control (SPC).

The SPC application tested was to identify abnormally long response times (outliers) in database management processes (Suh 1998). Hamburg's (1997) simulated training data set was also used in conjunction with the robust backpropagation algorithm in several of the trials. Robust backpropagation (especially when used with Hamburg's algorithm) was found to be extremely successful in this SPC application. Further applications are currently under study.