@INPROCEEDINGS{lang09ecml, author = "Lang, T. and Toussaint, M.", title = "Relevance Grounding for Planning in Relational Domains", booktitle = "Proc.\textasciitilde of the European Conf.\textasciitilde on Machine Learning (ECML)", year = "2009", month = "September", abstract = "Probabilistic relational models are an efficient way to learnand represent the dynamics in realistic environments consisting of manyobjects. Autonomous intelligent agents that ground this representationfor all objects need to plan in exponentially large state spaces and largesets of stochastic actions. A key insight for computational efficiency isthat successful planning typically involves only a small subset of relevantobjects. In this paper, we introduce a probabilistic model to representplanning with subsets of objects and provide a definition of object rele-vance. Our definition is sufficient to prove consistency between repeatedplanning in partially grounded models restricted to relevant objects andplanning in the fully grounded model. We propose an algorithm that exploits object relevance to plan efficiently in complex domains. Empiricalresults in a simulated 3D blocksworld with an articulated manipulatorand realistic physics prove the effectiveness of our approach.", folder = "Robotics", pdf = "http://www.user.tu-berlin.de/lang/pub/lang09ecml.pdf", pdfurl = "http://www.user.tu-berlin.de/lang/pub/lang09ecml.pdf" }