The effect of imperfect error detection on reliability assessment via life testing

Measurement of software reliability by life testing involves executing the software on large numbers of test cases and recording the results. The number of failures observed is used to bound the failure probability even if the number of failures observed is zero. Typical analyses assume that all failures that occur are observed, but, in practice, failures occur without being observed. In this paper, we examine the effect of imperfect error detection, i.e. the situation in which a failure of the software may not be observed. If a conventional analysis associated with life testing is used, the confidence in the bound on the failure probability is optimistic. Our results show that imperfect error detection does not necessarily limit the ability of life testing to bound the probability of failure to the very low values required in critical systems. However, we show that the confidence level associated with a bound on failure probability cannot necessarily be made as high as desired, unless very strong assumptions are made about the error detection mechanism. Such assumptions are unlikely to be met in practice, and so life testing is likely to be useful only for situations in which very high confidence levels are not required