Towards graphical partially observable Monte-Carlo planning

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.