Causally Deterministic Markov Decision Processes

Abstract

Probabilistic systems are often modeled using factored versions of Markov decision processes (MDPs), where the states are composed out of the local states of components and each transition involves only a small subset of the components. Concurrency arises naturally in such systems. Our goal is to exploit concurrency when analyzing factored MDPs (FMDPs). To do so, we first formulate FMDPs in a way that aids this goal and port several notions from concurrency theory to the probabilistic setting of MDPs. In particular, we provide a concurrent semantics for FMDPs based on the classical notion of event structures, thereby cleanly separating causality, concurrency, and conflicts that arise from stochastic choices. We further identify the subclass of causally deterministic FMDPs (CMDPs), where non-determinism arises solely due to concurrency. Using our event structure semantics, we show that in CMDPs, local reachability properties can be computed using a "greedy" strategy. Finally, we implement our ideas in a prototype and apply it to four models, confirming the potential for substantial improvements over state-of-the-art methods.