Abstract
Macro-operators are a common reformulation method in planning that adds high-level operators corresponding to a fixed sequence of primitive operators. We introduce meta-operators, which allow using different sequences of actions in each state. We show how to automatically verify whether a meta-operator is valid, i.e., the represented behavior is always doable. This can be checked at once for all instantiations of the meta-operator and all reachable states via a compilation into Stackelberg planning, a form of adversarial planning. Our results show that meta-operators learned for multiple domains can often express useful high-level behaviors very compactly, improving planners' performance.
Keywords
- Planning and Scheduling: PS: Planning algorithms
- Planning and Scheduling: PS: Learning in planning and scheduling
- Planning and Scheduling: PS: Search in planning and scheduling
