KIT | KIT-Bibliothek | Impressum | Datenschutz

Towards Graphical Partially Observable Monte-Carlo Planning : Technical Report IES-2015-09

Pfrommer, Julius


Sample-based online algorithms are state of the art for solving Partially Observable Markov Decision Problems (POMDP). But also the state of the art solver POMCP still suffers from the curse of dimensionality and curse of history. In Distributed POMDP, independent agents jointly optimise their actions under some coordination mechanism where every agent has access to a subset of the observations. In this work, we introduce Graphical POMDP (GPOMDP) drawing from existing Distributed POMDP appraoches as well as graph-based formulations as found in graphical probabilistic models. Further, we propose the Graphical POMCP (GPOMCP) algorithm that combines POMCP with message passing similar to the Belief Propagation (BP) algorithm from Graphical Probabilistic Models. In preliminary tests, GPOMCP shows good performance on a common Distributed POMDP benchmark.

Volltext §
DOI: 10.5445/KSP/1000054312
Cover der Publikation
Zugehörige Institution(en) am KIT Institut für Anthropomatik und Robotik (IAR)
Publikationstyp Proceedingsbeitrag
Publikationsjahr 2016
Sprache Englisch
Identifikator ISBN: 978-3-7315-0519-8
ISSN: 1863-6489
KITopen-ID: 1000060377
Erschienen in Proceedings of the 2015 Joint Workshop of Fraunhofer IOSB and Institute for Anthropomatics, Vision and Fusion Laboratory. Ed.: J. Beyerer
Verlag KIT Scientific Publishing
Seiten 113-125
Serie Karlsruher Schriften zur Anthropomatik / Lehrstuhl für Interaktive Echtzeitsysteme, Karlsruher Institut für Technologie ; Fraunhofer-Inst. für Optronik, Systemtechnik und Bildauswertung IOSB Karlsruhe ; 24
Relationen in KITopen
KIT – Die Forschungsuniversität in der Helmholtz-Gemeinschaft
KITopen Landing Page