FLTL: A Forward-Looking Approach

As noted in Section 1 above, the two key issues facing approaches to NMRDPs are how to specify the reward functions compactly and how to exploit this compact representation to automatically translate an NMRDP into an equivalent MDP amenable to the chosen solution method. Accordingly, our goals are to provide a reward function specification language and a translation that are adapted to anytime state-based solution methods. After a brief reminder of the relevant features of these methods, we consider these two goals in turn. We describe the syntax and semantics of the language, the notion of formula progression for the language which will form the basis of our translation, the translation itself, its properties, and its embedding into the solution method. We call our approach FLTL. We finish the section with a discussion of the features that distinguish FLTL from existing approaches.