Learning Probably Completable Plans
Abstract
In completable planning, a planning system is given the ability to defer goals which it can prove to be achievable. This has the advantages of allowing the use of runtime information in planning and enabling a planner to use less precise a priori information without sacrificing guarantees of success. In this paper, we extend completable planning to goals which are only probably achievable, thus extending its scope to a wider variety of problems. We also define completable plans in terms of its constituent reactive plan components, conditionals, and repeat-loops, which achieve the deferred goals, and we discuss the costs incurred by completable planning in terms of runtime evaluation cost, plan flexibility, a priori planning cost, and guarantee of success. In extending completable planning to probable achievability, we also introduce incremental explanation-based learning strategies for learning probably completable conditionals and probably completable repeat-loops, and demonstrate the learning of a probably completable plan in a simple train route-planning example.
Document Details
- Document Type
- Technical Report
- Publication Date
- Apr 01, 1991
- Accession Number
- ADA236853
Entities
People
- Gerald F. Dejong
- Melinda T. Gervasio
Organizations
- University of Illinois Urbana–Champaign