N-version programming -- the errors are in ourselves

[MellorPeter at aol.com]

Regarding the thread in RISKS-24.71 and 72, the results of Knight and
Leveson's famous N-version experiment show that, if any three of the
replicates from among those they had written were combined in a
two-out-of-three voting configuration, the resulting fault-tolerant system
would have a probability of failure 19 times smaller than one of the
replicates on its own.

This is not as much as fully independent failure would yield, but it is a
significant improvement.

Peter Mellor +44 (0)20 8459 7669   Mobile: 07914 045072  MellorPeter at aol.com

------------------------------