UC Berkeley

We address the problem of verifying Probabilistic Computation Tree Logic (PCTL) properties of Markov Decision Processes (MDPs) whose state transition probabilities are only known to lie within uncertainty sets. We first introduce the model of Convex-MDPs (CMDPs), i.e., MDPs with convex uncertainty sets. CMDPs generalize Interval-MDPs (IMDPs) by allowing also more expressive (convex) descriptions of uncertainty. Using results on strong duality for convex programs, we then present a PCTL verification algorithm for CMDPs, and prove that it runs in time polynomial in the size of a CMDP for a rich subclass of convex uncertainty models. This result allows us to lower the previously known algorithmic complexity upper bound for IMDPs from co-NP to PTIME. We demonstrate the practical effectiveness of the proposed approach by verifying a consensus protocol and a dynamic configuration protocol for IPv4 addresses. © 2013 Springer-Verlag.