Learning Hierarchies of Abstraction Spaces

Abstract

Hierarchical planning is an effective technique for reducing search in planning. Previous work on hierarchical planning has primarily focused on using abstraction spaces; the question of how the abstractions are formed remained largely unexplored. This paper describes Alpine, a system for learning abstraction spaces for use in hierarchical planning. Starting from only an axiomatization of the operators and example problems this system can learn detailed abstraction spaces for a domain. This is done using a theory of what makes a good abstraction space for hierarchical planning and then learning abstractions with the desired properties. The learned abstractions provide a significant performance improvement in PRODIGY, a domain-independent problem solver. The paper shows that Alpine can produce more detailed and effective abstractions using less knowledge than ABSTRIPS, a well-known system that partially automated the formation of abstraction spaces.