Browse Theses

Traditionally, abstraction in planning has been accomplished by either state abstraction or operator abstraction, neither of which has been fully automatic. We present a new method, predicate relaxation, for automatically performing state abstraction. Predicate relaxation generates abstraction hierarchies that, for some domains, can be more useful than those generated by previous abstraction mechanisms. PABLO, a nonlinear hierarchical planner, implements predicate relaxation. Theoretical, as well as empirical results are presented which demonstrate the potential advantages of using predicate relaxation in planning. Relaxed predicates can also be used by PABLO to achieve a limited form of reactivity, whereby an executable sequence of actions is constructed in case of interruption.

We also present a new definition of hierarchical operators that allows us to guarantee a limited form of completeness. This new definition is shown to be, in some ways, more flexible than previous definitions of hierarchical operators. The ability to plan using such operators has been incorporated into PABLO.

Finally, a Classical Truth Criterion is presented that is proven to be sound and complete for a planning formalism that is general enough to include most classical planning formalisms that are based on the STRIPS assumption.