A software component that is reused in diverse settings can experience diverse operational environments. Unfortunately, a change in the operating environment can also invalidate past experience about the component's quality of performance. Indeed, most statistical methods for estimating software quality assume that the operating environment remains the same. Specifically, the probability density governing the selection of program inputs is assumed to remain constant. However, intuition suggests that such a stringent requirement is unnecessary. If a component has been executed very many times in one environment without experiencing a failure, one would expect it to be relatively failure-free in other similar environments. This paper seeks to quantify that intuition. The question asked is, "how much can be said about a component's probability of failure in one environment after observing its operation in other environments?" Specifically, we develop bounds on the component's probability of failure in the new environment based on its past behavior.
展开▼
