Michael D. Moffitt
We explore a means to both model and reason about partial observability within the scope of constraint-based temporal reasoning. Prior studies of uncertainty in Temporal CSPs have required the realization of all exogenous processes to be made entirely visible to the agent. We relax this assumption and propose an extension to the Simple Temporal Problem with Uncertainty (STPU), one in which the executing agent is made aware of the occurrence of only a subset of uncontrollable events. We argue that such a formalism is needed to encode those complex environments whose external phenomena share a common, hidden source of temporal causality. After characterizing the levels of controllability in the resulting Partially Observable STPU and various special cases, we generalize a known family of reduction rules to account for this relaxation, introducing the properties of extended contingency and sufficient observability. We demonstrate that these modifications enable a polynomial filtering algorithm capable of determining a local form of dynamic controllability; however, we also show that there do remain some instances whose global controllability cannot yet be correctly identified by existing inference rules, leaving the true computational complexity of dynamic controllability an open problem for future research.
Subjects: 3.6 Temporal Reasoning; 15.2 Constraint Satisfaction
Submitted: Apr 24, 2007