Abstract:

Code inspection by human inspectors is considered to be the most
efficient method for detecting some kinds of faults ("bugs").
Currently available methods enable after the completion of the
inspection process to estimate the number of undetected faults. If
this number is unacceptable high, a further inspection is needed.
Another possibility is that the number of employed inspectors is
higher than needed, i.e. a waste of resources.  We developed a new
estimator, which enables ahead of the inspection process to estimate
the minimal number of inspectors required for a successful inspection.
Our statistical estimator is based on the observation that different
inspectors detect different sets of faults. A considerable overlap
between these sets of faults indicates that a large proportion of all
the faults have been detected. A small overlap indicates that there
are many undetected faults. The estimator assumes that the "ease of
detection" of the different faults is linearly distributed. Non linear
distributions have also been tried. The estimator has been tested with
both computer science students and industry engineers. It has been
tested for inspection of faults in code and for faults in user
requirements documents. The experiments suggest that the estimator can
provide sufficient accurate estimates of the minimal number of
required inspectors.