Experience with 'two version' programming

This paper summarizes the experience gained from reviewing a software program for attitude prediction of a satellite developed independently by two teams. In the review we found that both teams made similar errors that could bring to equivalent failures at the system level. The study showed that the cause of this situation could be traced back to the requirement phase and the physical model that was used. This physical model was based on an earlier one, developed for a system of a previous generation. Assumptions made for the previous system were adopted for the new one without updating them or examining their validity. Since both teams were exposed to the system history, these assumptions were found in both versions of the program. Moreover, comparing the output of the two versions increased confidence in both and led to a test plan that did not cover many cases that would have been otherwise analyzed. (author). 13 refs