|
SNN-1: Probabilistic Knowledge Representations |
|
Organization responsible: SNN Adaptive Intelligence People involved: Project description: This project aims to investigate, develop and evaluate methods for probabilistic reasoning in large graphical models for decision making.
Real world decision making often takes place in a complex, partially observable dynamic environment with inherent uncertainties. Graphical models provide a compact and mathematically consistent representation language of such an environment. The advantage of the graphical representation is that models are represented by local structures, while being globally consistent.
A problem with this approach is that probabilistic inference (state estimation, computation of optimal decisions, control) in large scale models is computational intractable. Therefore approximate methods are needed. During the last years, many novel approximation methods for graphical models have been developed. In particular, advances have been made in the so-called variational approaches, in which the global model is approximated by a collection of local quantities. Well known is the loopy belief propagation, which is an exact algorithm when the graph is a tree, but approximate with if the graph has loops. A powerful generalization of this method is the so-called Cluster Variation Method (CVM), which is currently considered as one of the most powerful approximations available.
In the context of this project, we will also study, develop and validate methods for probabilistic state estimation in real-world graphs and stochastic optimal control. The research aims at study and design of (approximate) algorithms to enable reasoning in systems to the order of 1000 variables, and to investigate the validity and applicability of these methods. These algorithms aim to deal with systems that are order of magnitude larger than currently possible. Publications:
|
