Probabilistic Model Checking and Reliability of Results

In formal verification, reliable results are of utmost importance. In model checking of digital systems, mainly incorrect implementations of the model checking algorithms due to logical errors are the source of wrong results. In probabilistic model checking, however, numerical instabilities are an additional source for inconsistent results. We motivate our investigations with an example, for which several state-of-the-art probabilistic model checking tools give completely wrong results due to inexact computations. We then analyze, at which points inaccuracies are introduced during the model checking process. We discuss first ideas how, in spite of these inaccuracies, reliable results can be obtained or at least the user be warned about potential correctness problems: (1) usage of exact (rational) arithmetic, (2) usage of interval arithmetic to obtain safe approximations of the actual probabilities, (3) provision of certificates which testify that the result is correct, and (4) integration of a "degree of belief" for each sub-formula into existing model checking tools.