Probabilistic Plan Verification through Acceptance Sampling

CIRCA is an architecture for real-time intelligent control. The CIRCA planner can generate plans that are guaranteed to maintain system safety, given certain timing constraints. To prove that its plans guarantee safety, CIRCA relies on formal verification methods. However, in many domains it is impossible to build 100% guaranteed safe plans, either because it requires more resources than available, or because the possibility of failure simply cannot be eliminated. By extending the CIRCA world model to allow for uncertainty in the form of probability distribution functions, we can instead generate plans that maintain system safety with high probability. This paper presents a procedure for probabilistic plan verification to ensure that heuristically-generated plans achieve the desired level of safety. Drawing from the theory of quality control, this approach aims to minimize verification effort while guaranteeing that at most a specified proportion of good plans are rejected and bad plans accepted.