Termination guarantees for CVI with global stopping criteria

Establish termination in finitely many steps of compositional value iteration with the proposed global stopping criteria—optimistic GSC and bottom-up GSC—for arbitrary input string diagrams of MDPs and weights, thereby providing a general termination guarantee.

Background

The paper introduces two sound global stopping criteria for compositional value iteration: an optimistic variant based on OVI and a bottom-up criterion using Pareto caches. These ensure error bounds upon termination.

However, the authors do not currently prove that their algorithm terminates; termination in undiscounted settings is subtle due to challenges like end components, which can arise compositionally.