Factorised contingency planning

Bram Ridder, Michael Cashmore, Maria Fox, Derek Long, Daniele Magazzeni

Research output: Chapter in Book/Report/Conference proceedingConference contribution book

Abstract

In this paper we consider one of the hardest problems in planning, contingency planning. Recent work has proposed translations for a specific class of contingency planning problems, characterised as Deterministic POMDPs, to classical planning problems. This class of contingency planning problems have deterministic actions and observations which makes it feasible to translate them into classical planning problems. This makes it possible to use mature classical planners like FF and Fast Downward to solve contingency planning problems. However, the translations proposed so far do not scale well and the results are not competitive with native contingency planners like POND and CLG. In this paper we improve upon previous translations by factorising the domain based on exploiting mutually independent observation actions. We show that our approach scales better compared to previous offline approaches in domains that are factorisable. For domains that do not factorise well we show that our approach is on-par with previous offline approaches.
Original languageEnglish
Title of host publicationProceedings of the 34th Workshop of the UK Planning and Scheduling Special Interest Group (PlanSIG 2016)
Place of PublicationHuddersfield
Number of pages8
Publication statusPublished - 16 Dec 2016

Keywords

  • Deterministic POMDPs
  • contingency planning
  • robotic systems
  • AI
  • artifical intelligence

Fingerprint Dive into the research topics of 'Factorised contingency planning'. Together they form a unique fingerprint.

  • Cite this

    Ridder, B., Cashmore, M., Fox, M., Long, D., & Magazzeni, D. (2016). Factorised contingency planning. In Proceedings of the 34th Workshop of the UK Planning and Scheduling Special Interest Group (PlanSIG 2016) Huddersfield.