Embedded Solution/Construction

Blind minimality is essentially the best achievable with anytime state-based solution methods which typically extend their envelope one step forward without looking deeper into the future. Our translation into a blind-minimal MDP can be trivially embedded in any of these solution methods. This results in an `on-line construction' of the MDP: the method entirely drives the construction of those parts of the MDP which it feels the need to explore, and leave the others implicit. If time is short, a suboptimal or even incomplete policy may be returned, but only a fraction of the state and expanded state spaces might be constructed. Note that the solution method should raise an exception as soon as one of the reward formulae progresses to $\mbox{$\bot$}$, i.e., as soon as an expanded state $\langle s, \phi\rangle$ is built such that $(\mbox{$\bot$}:r) \in \phi$, since this acts as a detector of unsuitable reward function specifications. To the extent enabled by blind minimality, our approach allows for a dynamic analysis of the reward formulae, much as in PLTLSTR [3]. Indeed, only the execution sequences feasible under a particular policy actually explored by the solution method contribute to the analysis of rewards for that policy. Specifically, the reward formulae generated by progression for a given policy are determined by the prefixes of the execution sequences feasible under this policy. This dynamic analysis is particularly useful, since relevance of reward formulae to particular policies (e.g. the optimal policy) cannot be detected a priori. The forward-chaining planner TLPlan [5] introduced the idea of using FLTL to specify domain-specific search control knowledge and formula progression to prune unpromising sequential plans (plans violating this knowledge) from deterministic search spaces. This has been shown to provide enormous time gains, leading TLPlan to win the 2002 planning competition hand-tailored track. Because our approach is based on progression, it provides an elegant way to exploit search control knowledge, yet in the context of decision-theoretic planning. Here this results in a dramatic reduction of the fraction of the MDP to be constructed and explored, and therefore in substantially better policies by the deadline. We achieve this as follows. We specify, via a $-free formula $c_0$, properties which we know must be verified by paths feasible under promising policies. Then we simply progress $c_0$ alongside the reward function specification, making e-states triples $\langle s,\phi,c \rangle$ where $c$ is a $-free formula obtained by progression. To prevent the solution method from applying an action that leads to the control knowledge being violated, the action applicability condition (item 3 in Definition 5) becomes: $a \in A'(\langle s,\phi,c\rangle)$ iff $a\in A(s)$ and $c\neq \mbox{$\bot$}$ (the other changes are straightforward). For instance, the effect of the control knowledge formula $\mbox{$\Box$}(p\rightarrow \raisebox{0.6mm}{$\scriptstyle \bigcirc$}
q)$ is to remove from consideration any feasible path in which $p$ is not followed by $q$. This is detected as soon as violation occurs, when the formula progresses to $\mbox{$\bot$}$. Although this paper focuses on non-Markovian rewards rather than dynamics, it should be noted that $-free formulae can also be used to express non-Markovian constraints on the system's dynamics, which can be incorporated in our approach exactly as we do for the control knowledge.
Sylvie Thiebaux 2006-01-20