Lifting the Limitations in a Rule-based Policy Language

The predicates that are used to encode a planning domain in PDDL often do not include concepts that are important for effectively reasoning about problems in the domain. In particular, the effectiveness of rule-based policies in a domain depend on the concepts that can be expressed in the language used to capture those policies. In this work we investigate complimenting planning domain descriptions with abstract concepts and methods for making distinctions between similar objects. We present an architecture that allows a rule-based policy to reason with these additional concepts, using them to reason over structures that the rules would not be able to reason over without support. We demonstrate that this is sufficient to allow a rule-based policy to provide control in benchmark domains with interesting structures and we argue that our architecture could allow control knowledge learners to learn policies that provide control in these domains.