Formal Multi-Objective Synthesis of Continuous-State MDPs

Abstract

This letter studies formal synthesis of control policies for continuous-state MDPs. In the quest to satisfy complex combinations of probabilistic temporal logic specifications, we derive a robust linear program for policy synthesis that is solved on a finite-state approximation of the system and is then refined back to a policy for the original system. This linear programming approach leverages occupation measures and enables the multi-objective optimizations needed to handle more complex probabilistic specifications.